Skip to main content
Glama

Part of the DEVIN ecosystem
Track: Build · Nature: product
For: AI engineers, Developers, Data Scientists
Interface: CLI / Python library / MCP server / Service
Path: AI engineers · step 3/3 — after devin-orchestrator

Where this fits

devin-memory

Renamed (Oct 2026): this repository moved from Icaro0310/devin-memory to Icaro0310/devin-brain — the PyPI package, CLI commands and MCP server name stay devin-memory; stars, issues and history are preserved by the GitHub redirect.

Unofficial community project. Not affiliated with, endorsed by, or sponsored by Cognition AI. "Devin" is a trademark of Cognition AI.

Linux · Personal Windows · Corporate Windows

Part of the awesome-devin ecosystem: the curated hub for the devin-* tools.

An anti-poisoning memory store for Devin: durable facts with provenance, versioning, and a quarantine gate — so agent memory can't be silently corrupted by a bad session or injected content.

Related MCP server: strata

The problem

Agent memory is a poisoning vector. Any tool that persists "facts" between sessions can be corrupted by a single bad session — an injected instruction or a pasted secret becomes a trusted belief in every future session, with no review step and no way to answer "where did this come from?".

Prior art

  • The Devin memory MCP (retain/recall/reflect over .devin/memory/memories.jsonl) — append-only, no screening, no session provenance. devin-memory exports to that exact line shape.

  • MemGPT / LangChain memory — persistence layers that optimize for recall, not for auditing or distrusting what was stored.

devin-memory adapts the memory-store idea; it adds the parts those tools don't have: a quarantine gate and provenance back to real session rows.

What makes it Devin-native

  1. Side-by-side: every entry can carry source_session_id + source_rowid, auditable against Devin's sessions.db via devin-internals' read-only store — the memory MCP cannot verify that a claimed source session (or a specific message row) ever existed. The quarantine gate also screens every write for secret and injection shapes.

  2. No-Devin: without sessions.db there is no session provenance to audit — the extra disappears.

  3. One sentence: it's a memory store that remembers where each memory came from — and quarantines suspicious ones until a human releases them.

Install

Python ≥ 3.10 required; install with uv (recommended) or pipx.

uv tool install 'devin-memory[mcp]'

or with pipx (alternative):

pipx install 'devin-memory[mcp]'

For development:

pip install -e ".[dev]"
pytest

Usage

# Store a fact (screened on write; suspect content lands in quarantine)
devin-memory retain "CI is green on Windows + Linux" --tags ci,status
devin-memory retain "..." --source-session <session-id> --source-rowid <n>
devin-memory retain "..." --workspace /path/to/project   # scope to a workspace
devin-memory retain "..." --expires-at 2026-12-31T00:00:00Z  # TTL: drops
                                   # out of recall/prime after this instant

# Keyword-ranked recall — returns active, unexpired entries only;
# scores are weighted by recency decay so fresh memories outrank stale ones
devin-memory recall "ci status" [--json] [--limit 5] [--tags a,b]
devin-memory recall "ci status" --include-expired  # also surface expired

# Quarantine lane: list, mark an existing entry, or release one
devin-memory quarantine                          # list with reasons
devin-memory quarantine <id> [--reason manual:x] # mark entry as quarantined
devin-memory quarantine <id> --hold-supersede    # mark + hold the supersession:
                                                 # the superseded parent returns
                                                 # to active until release/retract
devin-memory quarantine --release <id>           # human override -> active
                                                 # (applies a held supersession)
devin-memory quarantine --digest [--json]        # audit hash (read-only)

# Contradictions: a conflicting retain is linked, not overwritten
devin-memory conflicts [--json]     # (newer, older) pairs; resolve with
                                    # supersede / retract / quarantine <id>

# Mine a session for durable knowledge -> proposed entries (inactive
# until reviewed); extraction is heuristic — see "Limitations"
devin-memory extract <session-id> --sessions-db path/to/sessions.db
devin-memory extract --latest --sessions-db path/to/sessions.db [--auto-approve]
devin-memory list --status proposed   # review queue
devin-memory approve <id>             # proposed -> active

# Context block for a UserPromptSubmit hook — active entries only,
# filtered by workspace + machine profile, bounded by ~4 chars/token;
# also appends a generated ecosystem-state block (see below)
devin-memory prime [--workspace PATH] [--max-tokens N]
                   [--state-md PATH] [--max-chars N]

# Versioning and housekeeping
devin-memory supersede <id> "corrected fact" [--expires-at <iso>]
                                          # default: inherit old expiry
devin-memory retract <id>
devin-memory list [--status active|proposed|quarantined|retracted] [--json]
devin-memory compact [--json]   # dedup report (dry-run)
devin-memory compact --apply    # retract copies + vacuum the file

# Audit an entry's provenance: the write-time seal always runs; with
# --sessions-db it also cross-checks source_session_id/source_rowid
devin-memory verify <id> [--sessions-db path/to/sessions.db]

# Export active memories to a memory-MCP-compatible JSONL
devin-memory export --out memories.jsonl

MCP server

devin-memory is also a real MCP server (stdio) — the same retain/recall pipeline with the quarantine gate on every write, callable from Devin, Claude Desktop, Cursor or any MCP client:

pipx install "devin-memory-mcp"

The devin-memory-mcp console script also ships inside the base devin-memory dist, so a bare uv tool install devin-memory (no extra) leaves a script that exits with error: The MCP server needs the 'mcp' extra — recovery steps live in the Troubleshooting section of each platform guide (Linux, Personal Windows, Corporate Windows).

Client config:

{
  "mcpServers": {
    "devin-memory": {
      "command": "devin-memory-mcp",
      "args": ["--db", "/path/to/memory.db"]
    }
  }
}

The same stdio server registers in other MCP hosts:

Zed — context_servers in settings.json:

{
  "context_servers": {
    "devin-memory": {
      "command": { "path": "devin-memory-mcp", "args": ["--db", "/path/to/memory.db"] }
    }
  }
}

JetBrains — Settings → Tools → AI Assistant → Model Context Protocol → Add:

{ "mcpServers": { "devin-memory": { "command": "devin-memory-mcp", "args": ["--db", "/path/to/memory.db"] } } }

Antigravity — ~/.gemini/config/mcp_config.json:

{ "mcpServers": { "devin-memory": { "command": "devin-memory-mcp", "args": ["--db", "/path/to/memory.db"] } } }

Tools: retain, recall, screen (dry-run the gate, no write), list, retract, supersede, quarantine, release, approve, conflicts, prime, verify, extract, compact, quarantine_digest. The release and approve promotion overrides are only registered when the server runs with DEVIN_MEMORY_MCP_PROMOTE=1 — otherwise promotion stays human-only via the CLI (devin-memory quarantine --release <id> / devin-memory approve <id>). The same flag gates extract's auto_approve parameter and quarantine's hold_supersede (restoring a superseded parent is promotion-shaped), both forced off without the opt-in so an AI client cannot skip the review step. Every tool returns structured data or a {"error", "detail"} object — nothing raises through the transport. DEVIN_MEMORY_DB works as an alternative to --db.

VS Code: adapters/vscode/ ships an extension that registers this server with the editor's built-in MCP support (server definition provider, VS Code ≥1.102) — launch command, optional DEVIN_MEMORY_DB and workspace cwd are declared there; the promotion opt-in is pinned off so the editor-facing server keeps its non-promoting surface.

adapters/ is a self-contained Devin plugin root (adapters/.devin-plugin/plugin.json + adapters/skills/devin-brain/SKILL.md), installed with devin plugins install Icaro0310/devin-brain#adapters — it ships in the repo, not inside the wheel. The manifest launches the server through uvx --from 'devin-memory[mcp]' devin-memory-mcp; the promotion opt-in (DEVIN_MEMORY_MCP_PROMOTE) is never set, so the plugin-facing server keeps its non-promoting surface. To run against a checkout instead of PyPI, point a local manifest copy at uvx --from '.[mcp]' devin-memory-mcp.

Learn from sessions with devin-learning

This companion CLI extracts candidate lessons from a sessions.db and writes reviewable skill drafts. It does not install drafts into a workspace by default.

devin-learning extract --sessions-db path/to/sessions.db --out ./learning-drafts
devin-learning review --out ./learning-drafts

# After reviewing drafts, explicitly allow output to a live skill directory:
devin-learning extract --sessions-db path/to/sessions.db --out .devin/skills --apply

review is a dry-run unless --apply is given; review --apply moves rejected drafts under _rejected/. The extractor reads session contents, so keep its output private until reviewed.

Once drafts sit in a skills tree, catalog-bridge registers them into the devin-skill-catalog registry as proposed entries — closing the write→gate→promote loop:

# scans .devin/skills/learned-*/SKILL.md and registers each as `proposed`
devin-learning catalog-bridge .devin/skills

# then the catalog gates and promotes them (its own CLI):
devin-skill-catalog quarantine skill:learned-testing . --apply
devin-skill-catalog promote  skill:learned-testing --apply
devin-skill-catalog activate skill:learned-testing --apply

The registry defaults to <config-dir>/.devin-ecosystem/skill-catalog.json (resolved like devin-skill-catalog; DEVIN_CONFIG_DIR wins) — override with --registry. Entries carry metadata only (kind, name, state, sha256 of the SKILL.md, origin dir, empty history) — draft bodies are never copied. The bridge is deterministic and idempotent: re-scanning an unchanged draft changes nothing, an edited draft refreshes sha256, and items the catalog already moved past proposed are reported as skipped and left untouched. --dry-run prints the plan without writing. See docs/skill-catalog-bridge.md for the contract.

For a shareable summary there is weekly-report — the public surface of the loop. It emits weekly-YYYY-Www.md with aggregate counts only (lessons by theme, promoted learned-* skills, quarantine-pending buckets): no session text, ids or paths, so the file is safe to commit. It resumes from the devin-learning-weekly cursor, falling back to the last --days (default 7).

devin-learning weekly-report --sessions-db path/to/sessions.db --out ./docs/learning

Memory states

active · proposed (extracted, awaiting approve) · quarantined (screened or manually flagged, awaiting release) · retracted (withdrawn or superseded). Only active entries surface in recall/prime/export — quarantined content is never printed and never recalled.

A supersede whose replacement screens into quarantine is held: the superseded version stays active — and keeps surfacing in recall — until release applies the supersession or retract rejects it. The same hold works after the fact via quarantine <id> --hold-supersede, which returns the already-retired parent to active while the successor is reviewed. While a hold is pending the entry cannot be superseded again.

An active entry can also carry expires_at (ISO-8601, set at retain/ supersede or changed via set_expires_at): once the instant passes the entry is excluded from recall/prime — excluded, not deleted, so it stays visible in list/export --all for audit (recall --include-expired opts out). Entries without expires_at never expire. Separately, recall applies an exponential recency decay to keyword scores (half-life 30 days, ops.RECALL_DECAY_HALF_LIFE_DAYS, tunable per call via half_life_days) — fresher memories win close calls while clearly stronger stale hits still rank first.

Conflicts, extraction and prime (heuristics)

  • Conflicts — a retain that gives the opposite directive about the same normalized subject as an existing active entry is stored alongside it with a conflicts_with link (devin-memory conflicts). The heuristic compares a stop-word-stripped "subject key" plus affirmative/prohibitive polarity — it deliberately misses reworded contradictions rather than mislinking facts.

  • extract scans one session's message_nodes (read-only via devin-internals) for durable-knowledge signals — user corrections ("na verdade", "actually", "the right way"), preferences ("always", "never", "sempre", "nunca"), discovered commands (backticked known tools) and paths. Candidates are screened like any write: clean ones land proposed, suspect ones quarantined. --auto-approve skips the review step.

  • prime emits a compact # devin-memory: recalled context (heuristic) block sized for a prompt hook. Entries scoped with retain --workspace only prime inside that workspace; entries written under a different machine profile never prime (the profile defaults to corporate — fail-closed). The block is followed by an ## ecosystem state section rendered from a generated current-state.md — see Ecosystem state injection.

  • compact deduplicates: non-retracted entries sharing exact content + workspace + profile collapse to one survivor — the liveliest status wins (active > proposed > quarantined), ties keep the oldest row. The other copies become retracted (never deleted — history and reasons stay auditable) and the survivor inherits the union of tags from copies sharing its status — quarantined or proposed tags never bleed into an active entry. Default is a dry-run report (live entries N→M, per-group keep/retracted ids, truncated content sha256 — never content); --apply executes and runs VACUUM on the file.

  • quarantine --digest prints a deterministic audit fingerprint of the lane: entry count, reason buckets by taxonomy family (free-form manual reasons collapse to manual:*), and a sha256 over the sorted per-entry fingerprints sha256(content)+sorted(reasons). Keyed on content, not row ids — two stores holding the same quarantined facts digest identically, and any membership or reason change moves it. Always read-only.

The store is ./memory.db by default — override with --db or DEVIN_MEMORY_DB. It is the only store this tool writes to; Devin's sessions.db, acp-messages/*.db and state.vscdb are only ever read.

Ecosystem state injection

prime also appends a compact ## ecosystem state block rendered from a generated current-state.md — the generated-timestamp line plus the drift, service health and open PRs sections, capped at --max-chars (default 2000) and ending in a …(truncated) marker when cut. File search order: --state-md, then $DEVIN_ECOSYSTEM_STATE, then ~/devin/devin-ecosystem/devin-powerups/reports/current-state.md, then ./reports/current-state.md. Missing or unreadable state prints only state: unavailable (<reason>) — the exit code stays 0, so hooks and scheduled writes never break on absent state.

Two usage patterns:

  • Push — refresh a rules file the harness already injects, on a schedule (cron / systemd timer / Task Scheduler), so every session gets current state without remembering to pull:

    devin-memory prime > ~/.devin/rules/ecosystem-state-live.md
  • Pull-on-demand — call devin-memory prime (or with --state-md / --max-chars) inside a prompt or hook when fresh state matters more than a cached copy.

Works with Devin alone (Devin-only mode)

devin-memory keeps a local memory store (JSONL) with provenance tracking and a quarantine lane — no external memory service, no network calls. Both console scripts (devin-memory and devin-learning) run on your machine only.

Honest caveat: write-time screening is a heuristic, not a guarantee — suspect entries land in quarantine for human review, so keep that habit.

Platform support

The memory store uses an explicit local SQLite path and the session database is provided with --sessions-db; no platform-specific path is assumed. Windows and Linux are supported and covered by CI.

Limitations

  • Extraction is heuristic, and proposed by default. extract lifts keyword-shaped sentences from one session into a proposed review queue — nothing becomes active without approve (or --auto-approve). For a richer lesson pipeline see devin-learning.

  • The screen is a filter, not a guarantee. Pattern-based secret detection and injection heuristics have both false positives (→ quarantine, one command to release) and false negatives. Run dedicated scanners (gitleaks, devin-redact) too — this complements them.

  • Recall ranking is keyword-based, deterministic and documented — no embeddings or semantic search in M1.

  • Provenance is tamper-evident, not self-verifying. Every write seals the insert-frozen columns (content, source_*, created_at, machine/profile/workspace, chain position) into provenance_hash — verify recomputes it: VERIFIED_SEALED on a match, TAMPER_SUSPECT on any post-hoc edit (fails the audit even when the sessions.db half passes), UNSEALED on pre-seal rows. The seal proves the row wasn't modified after the fact; it does not prove the recorded source was truthful — a fabricated source_session_id seals cleanly. --sessions-db remains the cross-check for that claim.

  • A pending hold blocks re-superseding. An entry whose replacement is still quarantined can't be superseded again until the hold is resolved (release/retract), so two quarantined forks never compete over one parent.

  • PyPI distributions: devin-memory for the CLI and optional MCP extra; devin-memory-mcp is the standalone MCP-registry package.

When to use this

  • You persist agent memory between sessions and want it distrust-by-default: every write screened, suspect entries quarantined for human release.

  • You need to answer "where did this memory come from?" — entries carry source_session_id/source_rowid, auditable via verify.

  • You want memory versioning — supersede/retract keep a history instead of silent edits.

  • You want to stay compatible: export writes the Devin memory MCP's memories.jsonl line shape.

When NOT to use this

  • You need semantic recall — ranking is keyword-based, no embeddings.

  • You expect the screen to catch everything — it is a heuristic filter; run dedicated scanners (gitleaks, devin-redact) alongside.

  • You expect extract to read intent — it matches keyword signals and defaults to proposed precisely because heuristics err.

FAQ

How do I stop agent memory from being poisoned by a bad session? Use devin-memory retain instead of appending to a raw store. Every write is screened for secret and injection shapes — suspect entries land in quarantine and only become active after a human runs quarantine --release <id>.

Can devin-memory prove a memory came from a real session? Yes, via recorded provenance. retain --source-session <id> --source-rowid <n> stores the claimed origin, and devin-memory verify <id> --sessions-db <path> audits it read-only against Devin's actual sessions.db — a fabricated source is checkable, not silently trusted.

Does devin-memory replace the Devin memory MCP? It complements it. The MCP is append-only with no screening; devin-memory adds quarantine, provenance and versioning, and devin-memory export --out memories.jsonl produces the exact line shape the MCP reads.

License

MIT — see LICENSE.

Maintenance

ActivityMaintained
ResponsivenessUnresponsive

Related MCP Connectors

Related MCP Servers

  • F
    license
    Not graded
    quality
    B
    maintenance
    Enables trustworthy AI memory operations including authoritative memory creation, hybrid recall with provenance, evidence tracking, and safe dry-run-first forget and rebuild operations through a stateless MCP endpoint.
    2
    -
  • A
    license
    Not graded
    quality
    A
    maintenance
    Enables agents to read and write a durable, append-only memory with verified provenance, supporting search, record capture, and context preparation across sessions and machines.
    MIT
  • F
    license
    Not graded
    quality
    B
    maintenance
    Enables AI agents to store, search, and resolve shared memories with full provenance, conflict detection, and an audit trail across MCP-compatible clients.
    -
  • A
    license
    Not graded
    quality
    B
    maintenance
    Enables agents to govern the memory store they already run by gating what gets written, filtering recall by caller identity and memory validity, executing verifiable forgetting, and recording refusals and deletions in a hash-chained audit trail. It exposes these admission, authorization, and evidence controls over MCP so that credentials and personal data never persist and every deletion can be proven.
    1
    MIT