ai DeMemory

Public source and package-distribution repository for ai DeMemory, a local multi-LLM memory tool for Codex, Claude, Gemini, Obsidian, and future clients. Private memory lives in separately bound vaults and is never repository state.

Markdown is the canonical source of truth. SQLite FTS, exports, reports, and future vector indexes are generated from Markdown and can be rebuilt.

Status

• Published stable release: ai-dememory 2.0.0 on PyPI and GitHub Releases.

• Current public source version: 2.1.0rc1 on main. It remains unreleased until an immutable v2.1.0rc1 tag, the canonical release workflow, index verification, and explicit release authorization all exist.

• MCP protocol baseline: stable 2025-11-25, with 2024-11-05 accepted for older clients.

• Runtime boundary: Python 3.11+ remains authoritative; Node is not a headless runtime dependency. See docs/adr/0254-python-node-runtime-boundary.md.

• Transport: local MCP stdio plus optional local REST API.

• License: Apache-2.0; use, modification and redistribution are permitted under the terms in LICENSE.

• Repository workflow: Codex owns routine implementation and release readiness; independent read-only review precedes PR readiness, while merge, tag and publication require explicit user approval.

• Remote HTTP, OAuth, automatic durable writes, and vector search are out of scope for this release.

The active modernization and selective MemPalace adoption plan is documented in docs/public-modernization-roadmap.md. The progressive documentation source is in site/, governed by the documentation-site plan. Two isolated, security-guarded workflows validate the artifact and prepare manual GitHub Pages delivery. They have no push trigger: merging source cannot deploy, and Pages enablement plus an exact-main dispatch remain separate production operations.

Related MCP server: Mneme Memory MCP

Quick Start

Stable 2.0.0 from PyPI

Install the published tool, create a private vault, inspect its passive setup plan, and generate local client configuration:

pipx install ai-dememory
ai-dememory init ~/code/my-memory
cd ~/code/my-memory
ai-dememory doctor
ai-dememory index
ai-dememory setup plan --json
ai-dememory setup health --json
ai-dememory mcp-config --client codex
ai-dememory mcp-client-smoke

uv users can install the same tool with uv tool install ai-dememory.

Stable 2.0.0 does not include setup wizard, resource-intensity/model-policy selection, or generated MCP idle leases. Those capabilities belong to the unreleased 2.1.0rc1 source line below and must not be presented as PyPI behavior.

If you want a reusable private GitHub vault template repo instead of creating a single local vault, export the packaged vault template:

ai-dememory vault-template export ~/code/ai-dememory-vault-template

Review the exported files, push them to a separate private repository, then mark that repository as a GitHub template. Keep the tool distribution repo separate from private memory vault repos.

Unreleased 2.1.0rc1 source setup and optional onboarding

To evaluate the current development line instead of stable PyPI, install it directly from GitHub in a clean pipx environment, then use its preview-first wizard:

pipx install git+https://github.com/GonzaloTorreras/ai-dememory.git
ai-dememory init ~/code/my-memory
cd ~/code/my-memory
ai-dememory doctor
ai-dememory setup plan --intensity balanced --model-policy off --json
ai-dememory setup wizard
# Optional: create reviewed personal/project memory in a separate plan
ai-dememory onboard

The initial wizard is preview-first and configuration-only. It asks for one bounded operating envelope, prints a reviewable summary and exact fingerprint, then asks whether to apply that same in-memory plan to .ai-dememory.toml. Declining leaves setup incomplete and writes nothing. setup wizard --json returns the same passive operational preview without prompting. Personal values, preferences, recommendations, and project profiles are never requested or created by setup; add them only through the separate ai-dememory onboard preview/apply contract. Every JSON, stdin, input-file, and --dry-run apply is bound to the exact reviewed fingerprint.

balanced is the default. active is the maximum bounded profile, not an unlimited mode. Separately choose host-model policy off, advisory, or proposals. ai-dememory itself makes zero model and embedding calls in all profiles; advisory and proposals only authorize bounded work by an already active host agent, whose normal token usage still applies. No policy automatically promotes durable memory. Installation and the wizard do not create personal memory or install MCP configuration, hooks, provider imports, or scheduler jobs. After a confirmed interactive apply, the wizard prints the exact vault-local health, index, MCP-config, and optional onboarding commands to run next.

Run from the repository root. On Windows PowerShell, use py -3 if python3 is not available.

python3 scripts/ai_dememory.py doctor
python3 scripts/ai_dememory.py validate
python3 scripts/ai_dememory.py validate --json
python3 scripts/ai_dememory.py secret-scan
python3 scripts/ai_dememory.py index
python3 scripts/ai_dememory.py search ai-dememory --limit 3
python3 scripts/ai_dememory.py search ai-dememory --why
python3 scripts/ai_dememory.py context ai-dememory --budget 2000
python3 scripts/ai_dememory.py graph --json
python3 scripts/ai_dememory.py setup plan --json
python3 scripts/ai_dememory.py setup health --json
python3 scripts/ai_dememory.py recall-fixtures packet --limit 50 --pending-offset 50 --invalid-offset 50 --write-report
python3 scripts/ai_dememory.py providers detect
python3 scripts/ai_dememory.py capture markdown --path ./notes.md
python3 scripts/ai_dememory.py maintenance status
python3 scripts/ai_dememory.py schedule plan --json
python3 scripts/ai_dememory.py schedule setup --dry-run

search --why reports both numeric scoring components and matched evidence fields such as matched_terms, matched_fields, matched_tags, and matched_aliases. MCP memory.search returns the same explanation object. context reads optional [context] defaults from .ai-dememory.toml; explicit CLI flags and MCP arguments take precedence. Use context --why or MCP memory.context with explain_results=true to include ranking evidence in assembled Markdown context. For public-repository work, use an explicit query with context --public-only --no-working-memory or MCP memory.context with public_only=true and include_working_memory=false. The ceiling filters non-public memories before the result limit is applied, ranks only revalidated canonical public Markdown, ignores generated lifecycle state, does not read generated working state, and rejects public_only combined with auto. search --public-only, MCP memory.search(public_only=true), and MCP memory.get(public_only=true) expose the same public-sensitivity boundary for narrower recall. setup plan --json is read-only and includes generated_reports command arrays for optional recall review, recall review packet, manual acceptance, manual acceptance packet, hook capture review, and release evidence handoff artifacts. It also includes generated_archive_status command arrays for read-only recall and manual acceptance packet archive inspection, plus generated_archive_retention command arrays for previewing generated packet archive cleanup candidates without deleting files. setup health --json is read-only and combines validation status, context config status, manual acceptance readiness, recall review status, vector readiness, scheduler environment/status, provider readiness, maintenance preflight commands, generated artifact state and freshness, generated packet archive cleanup counts, lock state, and review queues into one local setup health summary. It reports separate core_ready, retrieval_evaluated, manual_maintenance_ready, automation_ready, maintenance_ready, integrations_ready, autonomy_ready, and release_ready dimensions. Automation is not ready until a namespaced scheduler receipt has been checked against the host after installation. The legacy ready field is deprecated and is only an alias for core_ready; it must not be interpreted as overall setup or release health. schedule plan --json is read-only and returns host scheduler commands, reviewed cron export entries, Docker command shapes when requested, and side-effect flags before any schedule setup command writes config or touches host scheduler state. Invalid resource overrides return resource_policy_valid=false with diagnostics and suppress every install/apply surface. maintenance status reports generated artifact state and freshness, generated packet archive cleanup counts, provider readiness, false-positive review due counts, stale suppression counts, conflict review counts, and hook capture review counts without reading provider chat files, deleting archives, or writing canonical memory.

Optional editable install from a local checkout:

python3 -m pip install -e .
ai-dememory doctor
ai-dememory search ai-dememory --limit 3

Use WSL paths for active Linux/web tooling checkouts when possible, but the repository tools are dependency-light and also run from native PowerShell.

Package installation is passive. Scheduled maintenance, provider imports, and Codex or Claude hook capture are explicit opt-in steps.

Architecture

• Markdown and Obsidian are the human-editable source of truth.

• Private GitHub syncs and versions canonical memory.

• SQLite FTS5 is the local retrieval and ranking layer.

• MCP exposes recall and write proposals to LLM tools.

• A local REST API exposes health, search, graph, reindex, and proposal endpoints for scripts and local dashboards that cannot launch MCP stdio.

• Graph generation uses indexes/memory.sqlite when available and falls back to Markdown parsing when the index is missing.

• Vector search is optional later, only if measured recall failures justify it.

See docs/architecture.md, docs/schema.md, docs/operations.md, docs/mcp-v2.md, and docs/mcp-v2-gap-analysis.md.

Distribution and user vault setup:

• Install guide: docs/install.md

• Local MCP server setup: docs/local-mcp.md

• Local REST API: docs/local-api.md

• Memory graph: docs/memory-graph.md

• Memory quality: docs/memory-quality.md

• Normative V3 roadmap: docs/v3-hybrid-visual-multiplatform-roadmap.md

• Machine-readable V3 frontier: contracts/planning/

• Historical R0-R7 research appendix, not an executable backlog: PLAN.md

• Shared memory governance roadmap: docs/shared-memory-governance-roadmap.md

• Import and capture: docs/import-capture.md

• Git lesson capture: docs/git-lessons.md

• Future vector migration: docs/vector-migration.md

• Operational loop: docs/operational-loop.md

• Review workflows: docs/review-workflows.md

• Sleep consolidation: docs/sleep-consolidation.md

• Scheduler and maintenance: docs/scheduler.md

• Scheduler/plugin blueprint: docs/scheduler-plugin-blueprint.md

• Local hook integrations: docs/hooks.md

• Codex plugin: docs/codex-plugin.md

• Distribution plan: docs/distribution.md

• Create a memory repo: docs/create-memory-repo.md

• GitHub vault template source: vault-template/

Safety Model

• Never store secrets, tokens, private keys, service-account JSON, cookies, recovery codes, or .env contents.

• Durable memories require human review before modification.

• LLMs may write proposals to inbox/llm-captures/, not directly to durable memory.

• Generated indexes live under indexes/ and can be rebuilt.

• Secret scanning and schema validation run before indexing.

• sensitivity: secret-prohibited is reserved for quarantined material and is rejected from canonical memory.

• private and sensitive memories are excluded from default search/MCP results and generated LLM context unless explicitly included by a local user.

• internal memory remains valid for a private vault but is not public-safe. Public-repository work must opt into the fail-closed public_only ceiling; default exclusion of only private/sensitive data is not an egress policy.

Repository Layout

• memories/durable/: reviewed durable values, preferences, policies, and facts.

• memories/active/: short-lived current working context.

• memories/projects/: project-specific memories and decisions.

• memories/tools/: tool-specific setup and behavior notes.

• inbox/: LLM proposals and raw captures awaiting human review.

• inbox/imports/: provider chat/session import candidates.

• inbox/git-lessons/: git history lesson candidates.

• inbox/session-events/: optional Codex/Claude hook metadata candidates.

• inbox/conflict-resolution/: reviewed conflict merge proposals.

• inbox/review-recommendations/: advisory LLM/client review recommendation artifacts that still require human action.

• inbox/sleep-consolidation/: generated sleep review packets.

• working/: generated current task state and handoffs.

• indexes/: generated SQLite and future vector indexes.

• distilled/: generated session context exports.

• reports/: generated review, scan, and consolidation reports.

• mcp/: MCP server skeleton and integration notes.

• scripts/: validation, scanning, indexing, search, export, and review tools.

• templates/: Obsidian-friendly memory templates.

MCP v2 Operation

• Client config examples: docs/mcp-client-config.md

• Protocol gap analysis: docs/mcp-v2-gap-analysis.md

• v2 release checklist: docs/release-v2-checklist.md

• PR-gated MCP runtime smoke: python3 scripts/ai_dememory.py mcp-smoke

• PR handoff: docs/pr-draft.md

• Roadmap status: docs/roadmap-status.md

• Normative V3 roadmap: docs/v3-hybrid-visual-multiplatform-roadmap.md

• Machine-readable V3 frontier: contracts/planning/

• Historical planning research, non-normative: PLAN.md

• Shared-memory governance roadmap: docs/shared-memory-governance-roadmap.md

Implemented MCP surface: 74 MCP tools.

Generated Codex, Claude, and generic private-vault configs expose only the four-tool server-enforced core profile by default. The checked-in public plugin is stricter: its three-tool public profile exposes only search, get, and context and forces public_only=true, include_sensitive=false, and no working-memory injection in the server. Additive working and review profiles are opt-in; admin preserves the complete server for backwards compatibility. Generated configs require an explicitly bound vault. See MCP tool profiles, including reproducible schema byte and estimated-token measurements.

• Tools: memory.search, memory.get, memory.write_proposal, memory.mark_seen, memory.reindex, memory.consolidate, memory.secret_scan, memory.graph, memory.doctor, memory.validate_status, memory.capture_miss, memory.recall_miss_candidate, memory.recall_fixture_status, memory.recall_review_plan, memory.recall_review_packet, memory.recall_review_packet_archive_status, memory.recall_review_packet_archive_retention_plan, memory.recall_miss_review, memory.vector_status, memory.roadmap_status, memory.context, memory.outcome, memory.lifecycle_scores, memory.maintenance_status, memory.import_chats, memory.capture_import, memory.git_lessons, memory.maintenance_run, memory.schedule_plan, memory.schedule_status, memory.schedule_environment, memory.hook_events, memory.hook_config, memory.hook_status, memory.hook_capture_review, memory.sleep_plan, memory.sleep_apply_reviewed, and memory.working_current, memory.working_status, memory.working_snapshot, memory.working_handoff, memory.providers_detect, memory.providers_status, memory.providers_plan, memory.setup_plan, memory.setup_health, memory.review_false_positives, memory.review_stale_false_positives, memory.false_positive_ignore, memory.false_positive_unignore, memory.review_conflicts, memory.conflict_dismiss, memory.conflict_keep, memory.conflict_merge_proposal, memory.review_modes, memory.review_configure_mode, memory.review_plan, memory.review_recommendation, memory.review_recommendations, memory.review_recommendation_archive_status, memory.review_recommendation_archive_restore_preview, memory.review_recommendation_outcome_report, memory.review_recommendation_outcome, memory.provenance_status, memory.acceptance_status, memory.acceptance_verify, memory.acceptance_plan, memory.acceptance_template, memory.acceptance_packet, memory.acceptance_packet_archive_status, memory.acceptance_packet_archive_retention_plan, memory.release_evidence, memory.release_evidence_report, and memory.publish_plan.

• Resources: memory://id/{id} and memory://path/{path} for public/internal canonical memories.

• Prompts: memory_recall_context, memory_capture_proposal, memory_review_inbox.

• Utilities: initialize, notifications/initialized, and ping.

The checked-in Codex plugin enables the server-forced three-tool public profile. Generated private-vault configs default to core; opt into working or review only for those workflows, and use the CLI or explicit admin profile when broad local execution tools are intended.

Safety defaults:

• MCP resources never expose private, sensitive, or secret-prohibited memories by default.

• Tools that can include sensitive content require an explicit include_sensitive argument.

• memory.write_proposal writes only to inbox/llm-captures/ and scans the rendered Markdown before writing.

• Working-memory tools write only generated operational state under working/; they do not promote durable memories.

• memory.secret_scan only accepts repository-relative paths through MCP.

• Review write tools only update .ai-dememory-ignore.toml or inbox/conflict-resolution/; false-positive suppressions report derived review_due and review_after_status fields from their review_after dates.

• Recall miss review writes only update reviewed frontmatter on files under inbox/recall-feedback/; fixture promotion remains a separate CLI review action.

• memory.mark_seen and memory.outcome return structured lifecycle receipts; outcome receipts report counters and metadata without echoing feedback notes.

• Docker is supported only for local stdio MCP usage with a bind-mounted vault; no ports or remote service are exposed.

Local REST API

Run a loopback-only API for local scripts and dashboards:

python3 scripts/ai_dememory.py api --host 127.0.0.1 --port 8765

Endpoints include /health, /search, /memories/{id}, /graph, /proposals, and /reindex. Graph responses are paginated. Non-loopback binds require both AI_DEMEMORY_API_KEY and TLS certificate/key arguments; there is no unauthenticated override. See docs/local-api.md.

Workflow

1. Capture new information as Markdown in inbox/ or the appropriate memories/ folder.

2. Run validation and secret scanning.

3. Rebuild the SQLite index.

4. Search or export context for LLM sessions.

5. Promote inbox proposals to durable/project/active memories only after review.

Validation And Release Gates

Run from the repository root:

python3 scripts/ai_dememory.py doctor
python3 scripts/ai_dememory.py verify-mcp
python3 scripts/ai_dememory.py ci-guard
python3 scripts/ai_dememory.py artifact-guard
python3 scripts/ai_dememory.py vault-setup-guard
python3 scripts/ai_dememory.py pr-template-guard
python3 scripts/ai_dememory.py pr-draft-guard
python3 scripts/ai_dememory.py acceptance-guard
python3 scripts/ai_dememory.py adr-guard
python3 scripts/ai_dememory.py release-checklist-guard
python3 scripts/ai_dememory.py release-check
python3 scripts/ai_dememory.py roadmap status --json
python3 scripts/ai_dememory.py api-smoke
python3 scripts/ai_dememory.py validate
python3 scripts/ai_dememory.py validate --json
python3 scripts/ai_dememory.py secret-scan
python3 scripts/ai_dememory.py eval-recall
python3 scripts/ai_dememory.py recall-fixtures status --json
python3 scripts/ai_dememory.py recall-fixtures review-plan --json
python3 scripts/ai_dememory.py recall-fixtures review-plan --write-report
python3 scripts/ai_dememory.py recall-fixtures packet --write-report
python3 scripts/ai_dememory.py recall-fixtures promote-miss --help
python3 scripts/ai_dememory.py recall-fixtures review-miss --help
python3 -m unittest discover -s tests -t .
python3 -m compileall -q scripts mcp/server ai_dememory_tool

CI runs artifact-guard before release gates and runs package-build-smoke --check-clean after install, package-build, and Docker smoke commands so stale package build metadata cannot be left behind by validation.

After a draft PR exists, run the runtime MCP smoke with the PR URL set:

AI_DEMEMORY_PR_URL="https://github.com/GonzaloTorreras/ai-dememory/pull/<number>" python3 scripts/ai_dememory.py release-check --strict
AI_DEMEMORY_PR_URL="https://github.com/GonzaloTorreras/ai-dememory/pull/<number>" python3 scripts/ai_dememory.py mcp-smoke

PowerShell equivalent:

$env:AI_DEMEMORY_PR_URL = "https://github.com/GonzaloTorreras/ai-dememory/pull/<number>"
py -3 scripts\ai_dememory.py release-check --strict
py -3 scripts\ai_dememory.py mcp-smoke

Generated artifact smoke commands:

python3 scripts/ai_dememory.py index
python3 scripts/ai_dememory.py search codex
python3 scripts/ai_dememory.py graph
python3 scripts/ai_dememory.py maintenance run --profile daily --dry-run --json
python3 scripts/ai_dememory.py maintenance run --profile daily
python3 scripts/ai_dememory.py maintenance run --profile weekly --dry-run --json
python3 scripts/ai_dememory.py maintenance run --profile weekly
python3 scripts/ai_dememory.py lifecycle scores --json
python3 scripts/ai_dememory.py lifecycle report
python3 scripts/ai_dememory.py sleep plan
python3 scripts/ai_dememory.py sleep --dry-run --json
python3 scripts/ai_dememory.py sleep --propose --id sleep_... --json
python3 scripts/ai_dememory.py sleep --apply-reviewed --id sleep_... --json
python3 scripts/ai_dememory.py sleep apply-reviewed --all
python3 scripts/ai_dememory.py working status --json
python3 scripts/ai_dememory.py hooks config --client codex
python3 scripts/ai_dememory.py hooks config --client claude
python3 scripts/ai_dememory.py hooks captures --json
python3 scripts/ai_dememory.py hooks captures --provider codex --review-status pending --json
python3 scripts/ai_dememory.py hooks captures --created-from 2026-06-01 --created-to 2026-06-30 --json
python3 scripts/ai_dememory.py hooks captures --write-report
python3 scripts/ai_dememory.py hooks review --help
python3 scripts/ai_dememory.py hooks archive --json
python3 scripts/ai_dememory.py hooks install --client all --dry-run
python3 scripts/ai_dememory.py providers configure codex --path "$HOME/.codex" --dry-run --json
python3 scripts/ai_dememory.py schedule plan --json
IMAGE_ID="$(docker image inspect --format '{{.Id}}' ai-dememory:local)"
python3 scripts/ai_dememory.py schedule plan --json --mode docker --image "$IMAGE_ID"
python3 scripts/ai_dememory.py schedule setup --dry-run --mode docker --image "$IMAGE_ID"
python3 scripts/ai_dememory.py schedule cron --json
python3 scripts/ai_dememory.py schedule doctor --json
python3 scripts/ai_dememory.py export-context
python3 scripts/ai_dememory.py consolidate --dry-run
python3 scripts/ai_dememory.py review false-positives
python3 scripts/ai_dememory.py review false-positives --due-only
python3 scripts/ai_dememory.py review stale-false-positives
python3 scripts/ai_dememory.py review conflicts
python3 scripts/ai_dememory.py review modes
python3 scripts/ai_dememory.py review plan --kind conflict
python3 scripts/ai_dememory.py review recommendation --kind conflict --target-id conf_example --recommendation collect_evidence --rationale "Needs human review." --recommended-by "Local LLM" --json
python3 scripts/ai_dememory.py review recommendations --json
python3 scripts/ai_dememory.py review recommendation-outcome --id rec_example --status accepted --reviewer "You" --reason "Reviewed." --json
python3 scripts/ai_dememory.py review recommendation-outcomes --json
python3 scripts/ai_dememory.py review recommendation-outcomes --limit 50 --offset 50 --invalid-offset 50 --json
python3 scripts/ai_dememory.py conflict resolve --id conf_example --keep mem_example --recommendation-id rec_example --reviewer "You"
python3 scripts/ai_dememory.py review recommendations-archive-status --limit 50 --offset 50 --invalid-offset 50 --json
python3 scripts/ai_dememory.py capture text --stdin --title "Session lesson"
python3 scripts/ai_dememory.py learn --git --days 7 --repo .
python3 scripts/ai_dememory.py learn --git --days 7 --repo . --write
python3 scripts/ai_dememory.py vector status
python3 scripts/ai_dememory.py recall-fixtures status --strict --max-age-days 14
python3 scripts/ai_dememory.py recall-fixtures review-plan
python3 scripts/ai_dememory.py recall-fixtures review-plan --write-report
python3 scripts/ai_dememory.py recall-fixtures packet --write-report
python3 scripts/ai_dememory.py recall-fixtures promote-miss --help
python3 scripts/ai_dememory.py recall-fixtures review-miss --help
python3 scripts/ai_dememory.py acceptance status
python3 scripts/ai_dememory.py acceptance plan
python3 scripts/ai_dememory.py acceptance plan --write-report
python3 scripts/ai_dememory.py acceptance packet --write-report
python3 scripts/ai_dememory.py acceptance packet --limit 50 --offset 50 --write-report

review modes and review plan include normalized [false_positives] and [conflicts] policy values from .ai-dememory.toml, including triage_policy, resolution_policy, scan toggles, and LLM auto-deny categories. These settings are exposed as review guidance; durable and canonical memory writes remain explicitly review-gated. Setting [false_positives].enabled = false or [conflicts].enabled = false makes the corresponding review reports and MCP listing tools return no candidates, and blocks review-state write commands for that workflow. JSON and MCP review listing responses include enabled and policy metadata for the relevant workflow. Generated Markdown review reports include the same compact Review Policy section, so archived false-positive, stale-suppression, and conflict reports show whether a workflow was disabled or simply empty. When [conflicts].scan_on_validate = true, validate also reports a non-blocking conflict review scan summary after frontmatter validation succeeds. When [conflicts].scan_on_consolidate = true, consolidate --dry-run includes the same non-blocking conflict review evidence in its generated report.

False-positive suppressions use [false_positives].review_after_days from .ai-dememory.toml when --review-after-days is omitted. New vaults default to 90 days, and explicit CLI/MCP arguments still override that policy per reviewed finding. Review state defaults to .ai-dememory-ignore.toml, or to [false_positives].ignore_file when configured inside the vault. Conflict reports and merge proposals use [conflicts].report_path and [conflicts].proposal_path when no explicit report path is supplied; both paths are constrained to the vault.

Product acceptance stays separate from automated package gates. This repository is operationally maintained by Codex: it prepares release PRs, versions, changelog entries and exact-artifact evidence. A release PR is not merged and no tag or package is published until the user explicitly authorizes those important actions. The manual tagger and publisher are separate exact-tuple dispatches. After both are authorized, the canonical publisher builds once, smokes the exact wheel and sdist, attests them, publishes through OIDC and verifies the installed index package. See docs/ai-operated-releases.md.

Manual acceptance remains useful for product and vault behavior, but it does not substitute for release authorization and is not consumed by the package workflow. Missing evidence is surfaced in the owner handoff for an explicit decision. After a reviewer uses a real MCP client, inspects an Obsidian vault, reviews a provider import, or verifies another checklist item, record that separate product evidence with:

ai-dememory acceptance record \
  --item mcp-client-installed \
  --reviewed-by "Reviewer Name" \
  --summary "Generated config was used with a real MCP client."
ai-dememory acceptance verify
ai-dememory release-evidence --write-report
ai-dememory release-evidence --strict
python scripts/ai_release_guard.py --version-only --json

The release-identity command succeeds only after an authorized release-prep change replaces the current version's Unreleased changelog heading with its actual release date.

publish-plan is read-only. Its compatibility workflow_url resolves the legacy hosted readiness preflight, never a publisher. That workflow requires confirm=preflight, a PR URL, read-only repository permission, and cannot request OIDC, target a package environment, upload artifacts, or publish. The plan retains the fields release_ready and publish_ready for API compatibility: TestPyPI local readiness may defer only the testpypi-publish acceptance item, while the local PyPI planner requires full release evidence. The canonical workflow cannot read private-vault receipts and does not enforce either field. Actual publication belongs exclusively to the manually dispatched .github/workflows/release.yml path for an exact immutable tag/commit tuple. Tag creation and publication are separate explicit owner authorizations after any remaining evidence gaps are disclosed; pushing a tag alone does not invoke the current publisher. testpypi-publish requires acceptance revision 3. Earlier or unversioned passing records from the former manual publisher or retired tag-push topology remain auditable but do not complete the current local acceptance/readiness signal.

Use ai-dememory release-evidence --write-report --report-path reports/v2-release-evidence.md when a handoff needs an explicit generated report target. The path must stay inside the memory root and the rendered Markdown is secret-scanned before writing. Add --reviewer "Reviewer Name" or set AI_DEMEMORY_REVIEWER when release evidence should pre-fill reviewer identity in embedded manual acceptance plans, templates, packets, and strict handoff commands. Add --pr-url https://github.com/... or set AI_DEMEMORY_PR_URL to carry the pull request URL into the same read-only handoff guidance. These fields do not record acceptance evidence.

Use ai-dememory acceptance plan to see remaining or blocked manual checks and the reviewed-evidence commands to run after each check. Each plan item also includes suggested_artifacts, such as a client log, reviewed inbox path, maintenance report, or TestPyPI workflow URL, so reviewers know what proof to attach before recording acceptance. Use ai-dememory acceptance plan --write-report to write the same read-only plan to reports/manual-acceptance-plan.md for handoffs. That generated report does not record evidence or count as acceptance; reviewers still need to run ai-dememory acceptance record with real proof. Use ai-dememory acceptance packet --write-report to write reports/manual-acceptance-packet.md, a reviewer-facing packet with fill-in sections, suggested artifacts, and pass/block record commands for every incomplete manual acceptance item. Use --limit and --offset to page large incomplete-item sections. Add --reviewer "Reviewer Name" and --pr-url https://github.com/... when a PR handoff should pre-fill reviewer and pull request context in the packet header. Add --archive when a review needs a timestamped copy under reports/manual-acceptance-packets/. It is still not evidence. Use ai-dememory acceptance packet-archive-status --json to list those generated packet snapshots with pagination metadata; the status command does not write files or record acceptance evidence. Use ai-dememory acceptance packet-archive-retention-plan --json to preview cleanup candidates after keeping the newest 30 generated packet snapshots by default; the retention plan does not delete files. Use ai-dememory acceptance template --item <item-id> when a reviewer needs a single-item evidence template without recording proof. The template is read-only guidance until ai-dememory acceptance record is run with reviewed details. Both acceptance plan and acceptance template accept --reviewer "Reviewer Name" and --pr-url https://github.com/... to pre-fill generated record commands with a reviewer and PR artifact while still leaving the item-specific summary for human review. ai-dememory release-evidence --json and the Markdown report also embed this manual acceptance plan so final handoffs include example record commands, suggested artifacts, and not only the remaining item descriptions. They also include release_blockers, a structured summary of dirty worktree, automated warning/failure, recall quality, and manual acceptance blockers that currently keep release_ready false. Recall freshness remains visible through recall_fixture_freshness and recall_fixture_review_plan; stale seed-only fixtures become a recall_fixture_review blocker only when there are pending or invalid recall miss files, recall eval is unavailable, or the current eval has failures. A clean current eval with no miss files stays visible as review evidence but does not force a synthetic miss before package release. Release evidence also includes vector_readiness, the same measured recall gate used by ai-dememory vector status, with creates_embeddings=false. If recall fixtures make a vector experiment eligible, release_blockers adds vector_readiness_review so the release handoff requires review before any embedding dependency or privacy model is approved. The same output includes a top-level next_actions list plus compact setup_health_summary and maintenance_summary objects so final handoffs show the ordered work remaining alongside validation, context defaults, scheduler readiness, hook capture review due counts, provider import readiness, recall review, vector readiness, generated artifact state and freshness, review queue counts, generated packet archive cleanup counts, and setup next actions without recording evidence, installing hooks, deleting archives, refreshing generated artifacts, or running maintenance commands. It also includes handoff_commands: copyable command arrays for writing release evidence, generating manual acceptance and recall review packets, running strict release evidence, verifying manual acceptance, checking recall freshness, evaluating TestPyPI/PyPI readiness, and running the publish guard. These commands are guidance; reviewers still need real evidence before recording acceptance or publishing. The handoff payload separates payload_* side-effect flags from command_side_effects; constructing release evidence remains read-only, while commands such as --write-report are explicitly marked as writing generated files and publish-plan commands are marked as running local read-only inspection if reviewers choose to run them. When reviewer and PR URL metadata are provided to release-evidence, the embedded acceptance_plan and acceptance_template handoff commands include that metadata so reviewers can copy the generated commands without replacing placeholders first. Weekly maintenance writes reports/sleep-plan.md as generated compaction review evidence; it does not write sleep review packets or mutate canonical memory. Use ai-dememory recall-fixtures check-miss --query "<query>" --expected-id <memory-id> --json before writing recall feedback. The check is read-only and reports whether the expected memory is outside the accepted rank window plus the exact capture-miss --dry-run and write commands to run if the miss is real. Use ai-dememory recall-fixtures packet --write-report when the weekly recall review needs a reviewer-facing handoff with fill-in fields, promote/reject commands, and final eval-recall and release-evidence --strict reminders. Add --reviewer "Reviewer Name" and --pr-url https://github.com/... when a PR handoff should pre-fill reviewer and pull request context in the packet header. Add --archive when the weekly quality review needs a timestamped copy under reports/recall-review-packets/. The packet is generated guidance only; it does not promote fixtures, close miss files, or write quality/recall-fixtures.json. Use ai-dememory recall-fixtures packet-archive-status --json to list generated recall packet snapshots with pagination metadata; the status command is read-only and does not promote fixtures. Use ai-dememory recall-fixtures packet-archive-retention-plan --json to preview cleanup candidates after keeping the newest 30 generated packet snapshots by default; the retention plan does not delete files.

If a manual check was attempted but cannot pass on the current workstation, record it as blocked instead of leaving the attempt invisible:

ai-dememory acceptance record \
  --item mcp-client-docker \
  --status blocked \
  --reviewed-by "Reviewer Name" \
  --summary "Docker is unavailable on this workstation."

Blocked records appear in release-evidence, but the item remains incomplete until a later passed record exists. ai-dememory acceptance verify exits nonzero until every manual acceptance item has reviewed passing evidence. ai-dememory release-evidence --strict also exits nonzero until automated evidence is clean and manual acceptance is complete.

Acceptance evidence is written under inbox/release-acceptance/ and secret-scanned before it is saved.

Direct script entry points remain available when debugging an individual tool:

python3 scripts/validate_memory.py
python3 scripts/validate_memory.py --json
python3 scripts/secret_scan.py
python3 scripts/index_memory.py
python3 scripts/search_memory.py codex
python3 scripts/export_context.py
python3 scripts/consolidate_memory.py --dry-run

MCP Server

Run as a stdio MCP server:

python3 scripts/ai_dememory.py mcp --stdio

The following lifecycle behavior describes the unreleased 2.1.0rc1 source line, not stable 2.0.0. The stdio server exits after 600 seconds without an MCP message by default. Generated setup configs use 120/600/1800 seconds for the minimal/balanced/active intensity profiles. This prevents completed agent sessions from retaining idle MCP processes indefinitely. Pass --idle-timeout-seconds 0 only when a separate supervisor owns cleanup. Package-owned Git, scheduler, install-smoke, and maintenance children receive closed stdin and run in an owned process group/tree. Every deadline reaps the complete descendant tree; MCP response reads also have a deadline and graceful EOF shutdown. Windows creates each child suspended, assigns it to a retained kill-on-close Job Object, and only then allows it to execute. POSIX uses an owned session/process group. Canonical scans, secret scans, graph pages, MCP input queues, captured output, and SQLite histories also have hard ceilings. These controls cover ai-dememory processes, not unrelated Node/Python tool servers owned by the host application.

PowerShell direct smoke examples:

'{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-11-25","capabilities":{}}}' | py -3 scripts\ai_dememory.py mcp --stdio
'{"jsonrpc":"2.0","id":2,"method":"ping"}' | py -3 scripts\ai_dememory.py mcp --stdio

Do not expose the stdio server as a network service without a separate authentication and authorization design.

CI

GitHub Actions runs compile, schema validation, secret scan, static MCP contract verification, release readiness, PR-gated strict release readiness and MCP runtime smoke on pull requests, unit tests, index rebuild, search smoke, package install smoke, package build smoke, Docker local MCP smoke, and the final package build artifact clean check. The PR-gated checks receive AI_DEMEMORY_PR_URL from the pull request event, and runtime smoke exercises a live stdio server process. The ordinary release readiness check runs before index generation; the strict PR-only release readiness check runs after index/search/recall smoke so doctor has generated index evidence.

Generated Artifacts

Generated artifacts must be reproducible from Markdown. SQLite databases, context exports, and reports are not canonical memory unless a human explicitly reviews and promotes their content into memories/. Before release, run ai-dememory artifact-guard or python3 scripts/ai_dememory.py artifact-guard to confirm generated outputs are not staged.