MEMO MCP Server
memo is a fully local, markdown-backed memory server that lets agents persist, retrieve, and maintain durable knowledge across sessions.
Save & store — persist facts/decisions with tags and types (
memo_save,memo_save_text), or offload bulky tool output/logs content-addressed (memo_offload).Search & recall — hybrid semantic+BM25 search, trace diagnostics, rerank, raw embeddings, and budgeted prompt-ready context packs (
memo_search,memo_context,memo_rerank,memo_embed_*).Answer questions — grounded, cited synthesis with conversational history (
memo_ask,memo_chat_ask,memo_evidence_packwith explicit abstention).Curate memories — get, update, rename, delete, forget/unforget, plus provenance and version rollback history (
memo_get,memo_update,memo_hash-style history/diff tools).Knowledge graph — explore entities, shortest/weighted paths, neighbors, communities, and code architecture via
memo_graph.Lifecycle & review — schedule/record reviews, invalidate or supersede facts over time for temporal validity (
memo_review_due,memo_mark_reviewed,memo_invalidate,memo_supersede).Operational journal — set focus, hand off work, raise/ack attention items, open conflicts, record task outcomes, and promote proven procedures (
memo_focus_*,memo_handoff_*,memo_attention_*,memo_conflict_*,memo_outcome_record,memo_procedure_*).Sessions & capture — start/track sessions, list them, run idle transcript capture, and pop pending notifications (
memo_start_session,memo_session_*,memo_idle_capture).Maintenance — rebuild indexes, detect near-duplicate clusters, and lint corpus issues (
memo_reindex,memo_consolidate,memo_lint).Diagnostics & extras — version/stats/embedder profile, live tool schema lookup, write-queue status, federation preview, watcher signals, journal hash-chain verification, and crush cache retrieval.

memo
Your coding agent starts every session with amnesia. memo fixes that — 100% on your own machine.
Persistent, searchable memory for Claude Code, Codex, Cursor, Cline, Devin, and OpenCode. No cloud, no API keys, no Ollama, no vector DB to run. And it spends fewer tokens, not more.

Install
curl -fsSL https://raw.githubusercontent.com/jagoff/memo/v4.16.0/install.sh | bashPrefer a package manager? uv tool install mlx-memo · pipx install mlx-memo · brew tap jagoff/memo && brew install mlx-memo
Then:
memo doctor # self-check
memo save 'we use Postgres, not Mongo' # save a decision
memo search 'what database did we pick?' # search by meaningThat's it. Your agents pick it up over MCP automatically — the installer wires every client it finds.
New Mac:
curl -fsSL https://raw.githubusercontent.com/jagoff/memo/v4.16.0/install.sh | bash
memo sync bootstrap git@github.com:yourname/memo-sync.gitAgent-managed setup:
curl -fsSL https://raw.githubusercontent.com/jagoff/memo/v4.16.0/install.sh | bash
memo doctor --strict-runtimeOn Linux or just want to look around first?
docker run --rm ghcr.io/jagoff/memo:latest memo doctorRelated MCP server: neuromcp
Why this saves you money
Most memory servers add context. memo is built to remove it.
Profile | Tools | Schema tokens |
| 43 | ~9.7k |
| 60 | ~13.2k |
| 165 | ~30.6k |
The default MCP surface is 43 tools, not 165 — 74% fewer tools, and about 68% less schema context: 43 tools / ~9.7k schema tokens versus 165 tools / ~30.6k tokens on the full surface — overhead paid every session, in every client.
Ambient recall injects one relevant memory before the model answers. The bundled Claude Code hook caps that injection at ~160 tokens. memo roi reports the real grounding and re-ask counts — the estimated-savings figure it used to print was removed in 4.14.0, because multiplying those counts by hardcoded constants was a savings claim memo could not support. For measured savings, memo tokens reads the provider's own usage counters through the context-compression proxy.
memo roi # value from grounded recalls and avoided re-asks
memo tokens # usage-savings ledgerThree things nothing else does
🕰️ Time-machine — query your knowledge as it was
memo as-of ask "what was the deploy strategy?" --date 2026-02-01
memo diff --from 2026-01-01 --to 2026-03-01Full historical reconstruction by reverse-replaying history.db. Useful when you need to know why past-you made a call, not just what past-you decided.
⚡ Contradiction radar — memory that notices when you change your mind
memo contradict scan # find conflicting facts corpus-wide
memo contradict triage # resolve: fuse / newer-wins / dismissChange a decision and memo flags the now-stale version, so the agent stops reintroducing what you already threw out.
🔮 Dream — it optimizes itself while you sleep
memo dream runA 7-phase nightly pipeline: inventory → mine signals → resolve conflicts → prune stale → synthesize cross-cluster insights → optimize → pre-warm the top-100 query embeddings so tomorrow's recall stays under 200 ms. Every run writes a receipt you can audit. Zero intervention.
How it works
Hybrid retrieval. A vector leg (MLX on Apple Silicon, sentence-transformers on CPU) and a BM25 leg (FTS5, diacritic-folding for Spanish) run in parallel, fuse via Reciprocal Rank Fusion, then go through an optional MLX cross-encoder rerank.
Markdown is the source of truth. Every memory is a plain .md file you can read, grep, and version-control. SQLite is a derived index that rebuilds from the files at any time — hand-edit in Obsidian and your edit wins on the next memo reindex. Nothing is locked in a database you can't open.
Prompts and memories stay on your machine. Embedder, reranker, and LLM all run in-process. No telemetry. Memory travels only if you point memo sync at a git remote you own. Normal startup is fully offline; remote update checks and auto-update require an explicit opt-in. → Privacy and network policy
Also in the box: cross-agent memo resume (reopen any session from any agent), cross-Mac git sync, a knowledge graph with optional codegraph symbol edges, encrypted secret storage, OCR/audio ingestion, evidence packs, outcome learning, signed federation, and a local chat UI over your memory (memo chat serve). → Full feature reference
How it compares
Verified July 2026 against each project's own docs. Corrections welcome — open an issue and I'll fix the table.
memo | ||||||
100% local, no cloud API | ✅ | ⚠️ | ⚠️ | ⚠️ | ✅ | ⚠️ |
Time-machine (rewind to any date) | ✅ | ❌ | ⚠️ | ❌ | ⚠️ | ⚠️ |
Contradiction detection + resolution | ✅ | ⚠️ | ⚠️ | ❌ | ❌ | ❌ |
Autonomous nightly maintenance | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ |
Token-economy MCP profiles | ✅ | ❌ | ❌ | ⚠️ | ✅ | ❌ |
Markdown / Obsidian as source of truth | ✅ | ❌ | ⚠️ | ❌ | ✅ | ❌ |
✅ first-class · ⚠️ partial, config-gated, or add-on · ❌ absent
Closest comparators are basic-memory (local-first + Obsidian + MCP — same thesis) and cipher (memory for coding agents).
Requirements
Support | |
macOS, Apple Silicon (M1–M4) | Full — MLX embedder + reranker + |
Linux / Ubuntu | Standalone CPU backend — search, recall, save. |
Intel Mac | Unsupported — current PyTorch releases do not ship Python 3.13 wheels for this platform |
Docker | Cross-platform, CPU backend · docs/docker.md |
Python ≥ 3.13 (the installer handles this via uv if you don't have it). First install pulls ~8 GB of models, 5–15 min. Optional: an Obsidian vault — without one, memo uses ~/Documents/memo/.
Docs
Install detail, installer knobs, new-Mac migration | |
Per-client MCP setup (Claude Desktop, Cursor, Cline, Continue) | |
Ambient recall, capture, and tuning | |
Full CLI reference (145 commands) + | |
All | |
Architecture and design notes | |
Privacy and network policy |
All 145 top-level CLI commands
Core: save search ask get edit rename delete list
Recall & Hooks: recall recall-hook context briefing continuity prewarm capture-tick capture-stop interject ask-gaps guard digest
Session & History: history as-of diff record-history session chat-session resume reflect mine-history episodes chronicle
Maintenance: reindex maintain review dream consolidate synthesize dedupe cross-dedup retier contradict coordinate terminal invalidate temporal compress-context ops
Analysis & Quality: health stats doctor journey-check lint drift analytics eval roi tokens token-savings usefulness gaps outcome profile confidence graduation hype definitive evidence
Knowledge Graph: graph entities entity extract-entities links version related
Advanced Search: embed rerank contextual retrieve context-pack chat chat-ask repo
Import / Export / Sync: import export backup restore sync ingest federation
Visualization: tui dashboard map logs hook-log
Setup & Config: init setup config install-mcp install-watcher uninstall-watcher install-slash install-statusline install-recall-hook install-shell-wrapper install-shims startup-banner migrate migrate-vault migrate-independence update upgrade self-update watch release onboard
Daemons: daemons recall-daemon ingest-daemon maint-daemon embed-daemon idle-daemon
Other: backend-native collaborative events feedback query mandate drift sleep-cycle operational ocr-image provenance secret verbatim mcp-command codex-badge debug-recall http-api proxy mine-git token-gate fix undo code-facts code-nudge code-health
Contributing
git clone https://github.com/jagoff/memo && cd memo
uv pip install -e '.[dev]'Issues and PRs welcome — see CONTRIBUTING.md. If memo is useful to you, a ⭐ genuinely helps other people find it.
MIT licensed. Built on Apple MLX, sqlite-vec, and codegraph.
Available Tools
60 toolsmemo_askARead-onlyIdempotent
Answer a question using memo retrieval and citations.
Read-only. Use when you want a synthesized answer grounded in durable
memories instead of raw hit lists. k, type, and snippet_chars tune
retrieval; source attributes consult logging. With client sampling
enabled, synthesis runs on the calling model (see synthesizer field).
Inputs: question (required), k, type, source, +3 more. Returns: object
memo_attention_ackAIdempotent
Acknowledge an attention item so it stops being surfaced.
Idempotent: returns {'acknowledged': false} when the id is unknown or already acknowledged; true only on the first acknowledgement.
Inputs: id (required), actor_id. Returns: object
memo_attention_addA
Add an item that must be surfaced to later agents.
Writes a journaled attention record; severities outside low|medium|high|critical are rejected.
Inputs: project (required), summary (required), severity. Returns: object
memo_chat_askARead-onlyIdempotent
Answer a conversational question with optional history and context.
Read-only. Use instead of memo_ask when prior turns or explicit
context should shape retrieval and synthesis. history is a list of
chat messages; session_id links the answer to a tracked memo session.
With client sampling enabled, synthesis runs on the calling model (see
synthesizer field).
Inputs: question (required), k, type, source, +4 more. Returns: object
memo_conflict_openA
Open a local, auditable reality conflict.
Writes a 'detected' conflict record to the hash-chained journal. Resolution is human-only — memo_conflict_resolve reports the CLI command a human must run.
Inputs: topic (required), summary (required), freeze_write, evidence_uris. Returns: object
memo_conflict_resolveARead-onlyIdempotent
Report the local human action required to resolve a conflict.
Read-only: never resolves anything itself — it returns the
memo operational conflict resolve command a human must run locally.
Inputs: id (required). Returns: object
memo_consolidateARead-onlyIdempotent
Detect near-duplicate clusters and propose merges.
Read-only — returns proposals without modifying the corpus.
Uses the AdvancedConsolidator under the hood (same as
memo_consolidate_list_archived). With client sampling enabled,
merge synthesis runs on the calling model up to
MEMO_SAMPLING_MAX_CALLS (see synthesizer field). max_clusters
bounds how many clusters come back; a single cluster's member list
is bounded separately by member_limit -- same-topic memories
cluster by design, so one cluster can hold most of the corpus.
Both defaults are derived from MEMO_MCP_RESPONSE_BUDGET_TOKENS rather
than picked for round numbers: a returned member carries a 600-char
body_preview, so the response costs roughly
max_clusters * member_limit * 850 chars and the previous 20 x 20
spent ~47k tokens against a 10k cap -- 20 clusters cannot be returned
with any useful member detail at all. 10 x 2 measures 3,978 tokens on
the conformance corpus and ~9,039 projected for a corpus whose bodies
actually reach the 600-char preview cap. Two sample members are what
it takes to SEE a duplicate; total still reports the cluster's real
size, and memo_get fetches any member by id. Raise either argument
to see more and accept a larger response (the budget middleware
refuses one that overruns rather than silently truncating it).
Inputs: type, threshold, max_clusters, member_limit. Returns: object
memo_contextARead-onlyIdempotent
Build prompt-ready memory context without calling the answer LLM.
Read-only. Unlike memo_search's raw hit list, this returns a budgeted, prompt-ready context pack for direct injection: static profile, dynamic recent context, query hits, omissions, and a readonly prompt wrapper.
Inputs: question (required), k, type, source, +4 more. Returns: object
memo_crush_retrieveARead-onlyIdempotent
Retrieve original content from crush cache.
When memo crushes a large JSON array during ingest, it offloads low-relevance rows to cache and embeds only the top-K rows. The context-compression proxy reuses the same cache to make a cut reversible. This tool recovers the original from either.
Inputs: hash_marker (required). Returns: object
memo_deleteADestructive
Permanently delete one memory by id or unique prefix.
Destructive and irreversible (no trash). Elicitation-capable clients are asked to confirm before the delete runs; other clients proceed unchanged. Resolves ambiguous short ids safely and returns an error instead of guessing. When cross-reference indexing is enabled, the response warns about memories that linked to the deleted record.
Inputs: id (required). Returns: object
memo_embed_batchARead-onlyIdempotent
Embed one or more document strings with memo's document embedder.
Read-only. Use for diagnostics or external indexing when you need document vectors from the same model memo uses internally. Pass a list of strings; an empty list returns no vectors without error.
Inputs: texts (required). Returns: object
memo_embed_queryARead-onlyIdempotent
Embed one query string with memo's query embedding path.
Read-only. Use for diagnostics or integrations that need the exact vector memo would use for retrieval queries. Rejects empty text and returns the vector, dimension, and model id.
Inputs: text (required). Returns: object
memo_evidence_packARead-onlyIdempotent
Return bounded, cited evidence for a question, or an explicit abstention.
Read-only, no side effects. Every snippet carries a memo:// citation;
when retrieval coverage stays under min_coverage the result is an
explicit abstention, never a fabricated answer.
Inputs: question (required), k, type, as_of, +2 more. Returns: object
memo_federation_previewARead-onlyIdempotent
Preview the exact memories an ACL would allow into a signed bundle.
Inputs: principal (required). Returns: object
memo_focus_clearAIdempotent
Clear a project's current focus.
Idempotent: returns {'cleared': false} when no focus was set.
Inputs: project (required), actor_id. Returns: object
memo_focus_setAIdempotent
Set the current focus for a project.
Replaces the project's previous focus; the change is journaled with the acting agent's identity. Safe to repeat with the same summary.
Inputs: project (required), summary (required), actor_id. Returns: object
memo_forgetADestructive
Mark one memory as forgotten without deleting its history.
Destructive in retrieval behavior: the memory is hidden from normal
recall/search surfaces until memo_unforget restores it. Pass reason
to record why the memory should no longer be surfaced.
Inputs: id (required), reason. Returns: object
memo_getARead-onlyIdempotent
Fetch one memory by id or unique id prefix.
Read-only. Returns the full memory record, None when it does not
exist, or an ambiguity error when the prefix matches multiple records.
Use memo_search or memo_list first when you do not know the id.
When MEMO_EMITTED_LEDGER is on and the resolved id already has an
entry in this session's emission ledger, this call counts as a
recovery against the feature's net-saving estimate (see
memo_cache_stats's emit_ledger.memo_get_after_digest) -- the
conservative rule the spec calls for, not proof this call actually
followed a digest pointer.
Inputs: id (required). Returns: result
memo_get_embedder_profileARead-onlyIdempotent
Return the active embedding model profile.
Read-only. Use this to inspect the model id, vector dimensions, normalization, and provider that memo uses for semantic search. Useful when verifying compatibility with stored vectors or external retrieval components.
Inputs: no named parameters. Returns: object
memo_graphARead-onlyIdempotent
Navigate the entity knowledge graph (read-only).
One consolidated explorer over memo's corpus graph. Pick a verb:
"path": shortest entity path fromatob(fewest hops)."why": weighted shortest patha->bas evidence — the same route with each hop's edge weight (how many memories bridge it), so a connection is explained, not just asserted."neighbors": direct neighbours ofentity(ora)."explore": a rich "what's around X" view ofentity(ora) — degree, neighbours, and the memories that mention it."communities": clusters of related entities (limitcaps count)."impact": changed code plus linked durable memories forcwd."architecture": a bounded CodeContextPack forcwd.entity(ora) selects a symbol/path focus;modeis scout, verify, or audit andcursorcontinues an incomplete page.
By default this navigates the MEMORY graph only (entities linked through
shared memories). Set include_code=True to also fold in the codegraph
code-structure layer (call/extends/etc. edges between code symbols).
Inputs: verb (required), a, b, cwd, +8 more. Returns: object
memo_handoff_consumeAIdempotent
Mark a handoff consumed so it stops surfacing to later agents.
Idempotent: returns {'consumed': false} when the id is unknown or the handoff was already consumed; true only on the first consume.
Inputs: id (required), actor_id. Returns: object
memo_handoff_createA
Create a durable handoff for another agent or session.
Writes a journaled handoff record that surfaces in memo_operational_state until some agent consumes it.
Inputs: project (required), summary (required), from_actor (required), to_actor. Returns: object
memo_historyARead-onlyIdempotent
List recent memory history events.
Read-only. Use for audit trails across the corpus or filter by op
and memory id when investigating a specific write, update, delete,
forget, or reindex action. Short ids are resolved safely.
Inputs: id, op, limit. Returns: result
memo_idle_captureA
Run idle capture on the current session.
Write tool. Extracts durable insights from the newest tracked session
transcript and saves them as memo records. Use after a meaningful unit
of work, or periodically for clients without Stop hooks. Do not use for
normal lookup; call memo_search, memo_ask, or memo_unified_briefing
when you only need to read existing memories. With dry_run=True, no
memories are written and the response only reports what would run.
Inputs: dry_run. Returns: object
memo_invalidateADestructive
Close one memory's truth-validity interval without deleting it.
Sets invalid_at in the record's frontmatter and index; the record itself is retained and stays readable.
Inputs: id (required), reason (required), at. Returns: object
memo_journal_verifyBRead-onlyIdempotent
Verify every native operational hash chain.
Inputs: no named parameters. Returns: object
memo_lintARead-onlyIdempotent
Inspect the memory corpus for maintenance issues.
Read-only. Returns grouped lint findings such as malformed metadata or
other records that may need cleanup, capped at limit per category —
the findings scale with the corpus and a whole report runs past any
client's response budget. Use memo_update, memo_delete, or vault edits
separately to fix findings.
Inputs: limit. Returns: object
memo_listARead-onlyIdempotent
List recent memories, optionally filtered by memory type.
Read-only. Use this to browse the corpus before choosing an id for
memo_get, memo_update, memo_rename, memo_delete, or history tools.
limit caps the number of returned records.
Inputs: type, limit. Returns: result
memo_mark_reviewedAIdempotent
Record review evidence, verify the memory, and schedule its next review.
Replaying the latest still-current review with identical evidence and actor is a no-op, so retries are idempotent.
Inputs: id (required), actor, evidence. Returns: object
memo_offloadAIdempotent
Offload a bulky payload (tool output, log, dump) out of the context
window: memo stores it content-addressed as a reference-tier memory
and returns {id, sha256, kind, synopsis, deduplicated, drill_down}.
Use memo_offload for bulk working-context dumps; use memo_save for
curated durable facts meant to be recalled. Idempotent per payload:
re-offloading identical content returns the existing id with
deduplicated: true instead of writing a new memory.
The synopsis is deterministic (no LLM): JSON keys, CSV headers, code
symbols, or compressed text. Reference tier is excluded from
auto-recall, so offloaded blobs never appear in the recall hook.
Fetch the full payload later with memo_get(id).
Inputs: content (required), title. Returns: object
memo_operational_stateARead-onlyIdempotent
Read current focus, handoffs, attention items, conflicts, and outcomes.
Read-only snapshot of the operational journal's current state. Returns
only what is still open unless include_closed is set, and only the
newest limit entries per section — open items awaiting human triage
accumulate faster than they settle, so an unbounded snapshot eventually
costs more context than the memories it describes.
Inputs: limit, project, include_closed. Returns: object
memo_outcome_recordAIdempotent
Record whether recalled memories helped a task succeed.
Feeds success/failure back into each cited memory's outcome stats —
the signal behind procedure promotion. Idempotent per
idempotency_key.
Inputs: status (required), task_id (required), memory_ids (required), idempotency_key (required), +2 more. Returns: object
memo_pop_notificationADestructive
Read and dismiss pending idle-capture notification.
Destructive only for the notification queue: returns the pending idle-capture message, then removes that transient notification so it is not shown twice. It never deletes memories or session transcripts. Use after memo_idle_capture or before replying to surface auto-saved insights; use memo_search or memo_list to read durable memory records.
Inputs: no named parameters. Returns: result
memo_procedure_candidatesARead-onlyIdempotent
List outcome-backed memories ready for procedural promotion.
Inputs: limit, min_utility, min_successes. Returns: object
memo_procedure_promoteA
Promote grounded memories into a reusable procedure/failure pattern.
Creates a new durable memory citing the sources; rejects sources whose
outcome stats don't meet the kind threshold.
Inputs: title (required), memory_ids (required), kind, reason, +2 more. Returns: object
memo_profileARead-onlyIdempotent
Return bounded stable and active memory with evidence metadata.
Inputs: cwd, limit, scope, source, +1 more. Returns: object
memo_provenanceARead-onlyIdempotent
Return provenance metadata for one memory.
Read-only. Use with a full id or unique prefix when you need origin, lineage, file path, or related audit details for a record before trusting, editing, or citing it.
Inputs: id (required). Returns: result
memo_record_diffARead-onlyIdempotent
Return recent history events for one memory in chronological order.
Read-only. Use to inspect how a memory changed over time before
editing or rolling back. Accepts a full id or unique prefix; limit
caps returned events and has_more reports truncation.
Inputs: id (required), limit. Returns: object
memo_reindexAIdempotent
Rebuild memo's searchable index from the markdown vault.
Writes only derived index state; markdown remains the source of truth.
Use after hand-editing vault files or changing indexing behavior.
force reprocesses records even if memo thinks they are current.
errors counts files that FAILED to index (parse error, embed failure,
refused path) — a subset of skipped, which also counts deliberate
skips (archives, chronicle, secrets). A non-zero errors means the
index is incomplete: fix the cause and re-run.
Inputs: force. Returns: object
memo_renameADestructive
Rename one memory title without changing its body or tags.
Destructive metadata edit. Use after memo_save or memo_search when a
record has the right content but the wrong title. Pass id for a
specific memory; omit it only immediately after a save, when memo can
target the most recent local save. Use memo_update instead when you
need to edit content, type, or tags.
Inputs: title (required), id. Returns: result
memo_rerankARead-onlyIdempotent
Rerank candidate memory hits for a query.
Read-only. Performs no retrieval of its own (unlike memo_search) — it only reorders caller-supplied hit dictionaries, e.g. after memo_search or another retrieval source, when you need the most relevant subset ordered for answer synthesis. When reranking is disabled in this install, hits pass through in input order without scores.
Inputs: hits (required), query (required), top_n, body_chars. Returns: result
memo_review_dueARead-onlyIdempotent
List records whose explicit review date passed or have an open conflict.
Inputs: limit, project. Returns: object
memo_saveA
Persist content to memo.
Use memo_save for a durable curated fact, decision, or preference; use memo_offload for bulk dumps of working context.
When extract is true (defaults to the MEMO_SAVE_EXTRACT flag, off),
the helper LLM decomposes content into atomic facts and saves each as
its own memory (mem0 ADD-model) instead of one opaque blob; tags
propagate to every fact. Returns an extraction summary
(status, saved ids, saved_titles, counts) rather than a single
record. If nothing extractable is found, the blob is saved verbatim.
Normal saves add action (created, corroborated, or revised) and
index_pending so callers can distinguish evidence from a new record.
scope controls the auto project:<repo> tag for THIS call only:
"global" skips it (the memory lands untagged → the global recall
tier, +0.10 boost everywhere); "project" or None keep the default
auto-detection. An explicit project: tag in tags always wins
either way.
defer_embed mirrors CLI memo save --defer-embed: it persists
markdown + BM25 immediately and marks the semantic vector pending for
memo_reindex. Extraction mode ignores it, matching the CLI.
Inputs: content (required), tags, type, extra, +5 more. Returns: object
memo_save_textA
Save a memory from text.
Write tool. Persists one plain note from text and returns the saved
memory id. Use for quick client-agnostic saves when you already have
the exact content to remember. Prefer memo_save when you need tags,
type, extraction, conflict handling options, or project/global scope.
Prefer memo_idle_capture for transcript-derived session insights.
title is optional; when omitted memo uses the first text line.
Inputs: text (required), title. Returns: object
memo_searchARead-onlyIdempotent
Search durable memories by text, vector similarity, or hybrid mode.
Read-only. Returns raw ranked hits — source records with ids, dates, tags, and excerpts. Use memo_context instead for a budgeted prompt-ready pack, memo_search_trace for scoring diagnostics, and memo_rerank to reorder hits you already have.
Inputs: query (required), file, mode, type, +7 more. Returns: object
memo_search_traceARead-onlyIdempotent
Search memories and include retrieval trace diagnostics.
Read-only. Debug variant of memo_search: use it when investigating ranking, filters, or recall misses rather than for normal lookup. Returns the same style of hits as memo_search plus trace metadata that explains how candidates were selected and scored at each stage.
Inputs: query (required), mode, type, limit, +2 more. Returns: object
memo_session_getARead-onlyIdempotent
Fetch metadata for one tracked memo session.
Read-only. Use after memo_session_list or memo_start_session when you need the stored transcript path, project, checkpoints, or other session details. Returns None when the session id is unknown.
Inputs: session_id (required). Returns: result
memo_session_listARead-onlyIdempotent
List tracked memo sessions.
Read-only. Use to find recent session ids, transcript paths, and
project context for capture or inspection. project narrows results
and limit caps the number of sessions returned.
Inputs: limit, project. Returns: result
memo_signal_listBRead-onlyIdempotent
List durable watcher markers, newest epoch first.
Inputs: limit, marker, min_epoch. Returns: object
memo_signal_rememberCIdempotent
Remember a durable, idempotent operational watcher marker.
Inputs: marker (required), epoch, fence, payload, +1 more. Returns: object
memo_start_sessionA
Start a new session for this client.
Write tool. Creates or refreshes a tracked session checkpoint used by
capture, grounding, and transcript lookup. Call once at the beginning
of a task or conversation. Pass session_id to continue a known
session; omit it to create one. Pass cwd to bind the session to a
project directory; omit it to use the current process directory.
Do not use this to save memories directly; call memo_save or
memo_save_text for durable content.
Inputs: cwd, session_id. Returns: object
memo_statsARead-onlyIdempotent
Return local memo corpus and runtime statistics.
Read-only. Use for diagnostics, health checks, and environment inspection. Includes corpus count, storage paths, embedder model, history error count, and recall daemon health when available.
Inputs: no named parameters. Returns: object
memo_supersedeADestructive
Close the old memory at the successor's validity start.
Only the old record is modified; the close time never precedes the old record's own validity start.
Inputs: new_id (required), old_id (required), reason (required). Returns: object
memo_terminal_listARead-onlyIdempotent
List deliverable terminals; empty while legacy TTY input is disabled.
Inputs: no named parameters. Returns: object
memo_tool_docsARead-onlyIdempotent
Look up a tool's live description and parameter schema by name.
Use this when a tool you expect (from a prior session, or from documentation) is missing from your current tool list: memo's proxy prunes schemas of recently-unused tools from context to save tokens — memo's own tools stay registered and callable regardless; other servers' tools are still reachable exactly as they were before pruning. Call this with the tool's name to get back what its full schema would have shown, then call the tool directly by name.
Inputs: name (required). Returns: object
memo_unforgetAIdempotent
Restore a previously forgotten memory to normal retrieval.
Idempotent write. Accepts a full id or unique prefix and returns whether a matching forgotten record was restored. Use after deciding a memory should participate in search and recall again.
Inputs: id (required). Returns: object
memo_unified_briefingARead-onlyIdempotent
Load a compact startup briefing from Memo's durable and operational state.
Read-only. Call before deciding or answering so prior durable facts can
ground the task. Pass cwd to bias project context and source to
attribute consult logs to the calling client.
Inputs: cwd, source. Returns: object
memo_updateADestructive
Patch fields on a memory. content replaces the whole body;
replace_old+replace_new is a surgical exact-string edit (old must
occur exactly once — unchanged text stays byte-identical); append
adds a paragraph. All three are versioned (memo_version_rollback).
Inputs: id (required), tags, type, title, +4 more. Returns: result
memo_versionARead-onlyIdempotent
Get memo version info — the package version and the backend protocol version. Use this to check which version of memo is running.
Inputs: no named parameters. Returns: object
memo_write_queue_statusARead-onlyIdempotent
Return process-local MCP write queue depth, waits, and rejections.
Inputs: no named parameters. Returns: object
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 60 tools
Many tools are clearly distinct, but there is real overlap in the save and retrieval clusters: memo_save vs memo_save_text vs memo_offload, and memo_search vs memo_context vs memo_ask vs memo_evidence_pack all require careful reading of the descriptions to choose correctly. The long descriptions help, but the boundaries are not obvious from names alone.
Every tool shares the memo_ prefix and uses readable snake_case, but the verb/noun order is inconsistent: core tools use verb_noun (memo_list, memo_search, memo_save_text) while operational tools often use noun_verb (memo_handoff_create, memo_focus_set, memo_attention_add). The naming is still readable, but the mixed conventions prevent a fully predictable pattern.
With 60 tools, this is well beyond the 25+ threshold and into the extreme range. While the domain is broad, many tools are micro-variants or internal plumbing that could be consolidated, and the sheer count will overwhelm an agent's tool-selection surface.
The tool surface is unusually thorough: full memory CRUD, search variants, retrieval augmentation, embeddings, history, sessions, maintenance, and an extensive operational journal with focus, handoffs, attention, conflicts, outcomes, procedures, and review lifecycles. There are no obvious dead ends for the stated memory-and-operational-state purpose.
Maintenance
Related MCP Connectors
Persistent memory for AI agents. Semantic search, memory graph, W3C DID identity.
Persistent, portable memory for AI assistants — your private memory graph, from any MCP client.
Persistent memory and knowledge graphs for AI agents. Hybrid search, context checkpoints, and more.
Persistent memory, hybrid search and a goal graph for AI agents, over stdio or remote HTTP.
Related MCP Servers
- AlicenseAqualityBmaintenanceBasic Memory is a knowledge management system that allows you to build a persistent semantic graph from conversations with AI assistants. All knowledge is stored in standard Markdown files on your computer, giving you full control and ownership of your data. Integrates directly with Obsidan.md21210,713 PyPI4,572AGPL 3.0
- AlicenseNot gradedqualityBmaintenanceSemantic memory for AI agents — local-first MCP server with hybrid search, knowledge graph, contradiction detection, and plan-then-commit consolidation.199 npm4AGPL 3.0
- AlicenseNot gradedqualityDmaintenanceSelf-hosted semantic memory for AI agents. Save worklogs, decisions, and notes via MCP, then recall them across sessions by meaning rather than keyword. Backed by Postgres + pgvector with local embeddings (multilingual-e5-base).1MIT
- FlicenseNot gradedqualityDmaintenancePersistent semantic memory MCP server for AI agents with hybrid search, LLM scoring, and decay engine, fully local.2-