Local Memory

Part of the StudioMeyer MCP Stack — Built in Mallorca 🌴 · ⭐ if you use it

local-memory-mcp

Persistent local memory for Claude, Cursor & Codex. 25 tools. Hybrid retrieval (BM25 + vector cosine, RRF). Bi-temporal asOf queries + fact supersession. LLM-free contradiction detection + reflection. Portable JSON export/import. Multilingual embeddings. No cloud. No API keys.

Your AI assistant forgets everything when you close the chat. This fixes that.

Learnings, decisions, people, projects — stored in a single SQLite file on your machine that never leaves your computer. Built-in Knowledge Graph, duplicate detection, FTS5 keyword search, and (new in v2) hybrid retrieval that fuses BM25 with on-device vector cosine via Reciprocal Rank Fusion. The embedding model is multilingual (DE / EN / ES / 100+ languages) and runs locally — no API keys, no cloud.

📄 Deep dive: WHITEPAPER.md — architecture, design principles, the local-first contract, and an honest comparison vs. other memory systems.

Not affiliated with danieleugenewilliams/local-memory-releases — that is a different "Local Memory" project with the same descriptive name. This package is published as @studiomeyer/local-memory-mcp — always use the scoped name to disambiguate.

A note from us

We have been building tools and systems for ourselves for the past two years. The fact that this repo is small and has few stars is not because it is new. It is because we only just decided to share what we have built. It is not a fresh experiment, it is a long story with a recent commit.

We love building things and sharing them. We do not love social media tactics, growth hacks, or chasing stars and followers. So this repo is small. The code is real, it gets used, issues get answered. Judge for yourself.

If it helps you, sharing, testing, and feedback help us. If it could be better, an issue is more useful. If you build something with it, tell us at [email protected]. That genuinely makes our day.

From a small studio in Palma de Mallorca.

Quick Start

Claude Code

claude mcp add memory -- npx -y @studiomeyer/local-memory-mcp

Claude Desktop

With Node.js 22 installed, add the following to claude_desktop_config.json. This uses the published npm package.

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@studiomeyer/local-memory-mcp"]
    }
  }
}

Cursor / VS Code

Add to .cursor/mcp.json or .vscode/mcp.json:

{
  "mcpServers": {
    "memory": {
      "command": "npx",
      "args": ["-y", "@studiomeyer/local-memory-mcp"]
    }
  }
}

Codex

# ~/.codex/config.toml
[mcp_servers.memory]
command = "npx"
args = ["-y", "@studiomeyer/local-memory-mcp"]

Automatic session tracking

You can make session tracking fully automatic so you never have to think about it.

Claude Code (CLAUDE.md): Add this line to your project's CLAUDE.md:

Always call memory_session_start at the beginning of each conversation and memory_session_end when done.

Claude Code (Hook): For a system-wide setup, add a SessionStart hook in ~/.claude/settings.json:

{
  "hooks": {
    "SessionStart": [{
      "hooks": [{
        "type": "command",
        "command": "echo '{\"hookSpecificOutput\":{\"hookEventName\":\"SessionStart\",\"additionalContext\":\"Call memory_session_start now.\"}}'",
        "timeout": 5
      }]
    }]
  }
}

Both approaches make Claude call memory_session_start automatically. The CLAUDE.md way is simpler, the hook way works across all projects.

What it does

When you start a conversation, the server loads context from your last sessions so the AI knows what you were working on.

During the conversation, the AI stores patterns, insights, and mistakes via memory_learn. It records facts about people, projects, and tools via memory_entity_observe — building a knowledge graph over time. Every stored row is also embedded into a local 384-dim vector via the multilingual-e5-small model.

When you search, the unified memory_search runs hybrid retrieval: FTS5 with BM25 ranking is fused with vector cosine via Reciprocal Rank Fusion (RRF, k=60). That bridges vocabulary mismatches ("send" finds "publish"), works across DE / EN / ES / 100+ languages, and matches even when the query has no exact token overlap with the stored content. If the vector extension can't load on your machine, search transparently falls back to FTS5-only — nothing breaks, you just lose the semantic half.

The duplicate gatekeeper still prevents storing the same information twice.

Hybrid Search (v2.0.0+)

memory_search({ query: "...", mode: "hybrid" })   // default
memory_search({ query: "...", mode: "fts" })       // keyword only
memory_search({ query: "...", mode: "vector" })    // cosine only

Architecture

• search_fts (FTS5, BM25) — keyword recall, the v1 path.
• embeddings (sqlite-vec vec0 virtual table, float[384]) — vector recall.
• Reciprocal Rank Fusion (k=60) combines the two when mode: "hybrid".
• Embeddings come from Xenova/multilingual-e5-small (Apache-2.0) via Transformers.js, q8-quantized (~30 MB cache). Model loads lazily on the first embed call; runs entirely on CPU.
• Auto-embed-on-insert covers learnings, decisions, and entity observations. Entities themselves are not embedded — their attached observations carry the semantic surface.

Multilingual. The default model is trained on 100+ languages with strong DE/EN/ES retrieval. Mixing languages in your stored data is fine — query in one language and the cosine half still surfaces relevant results in another.

Environment overrides

• MEMORY_EMBED_DISABLED=1 — force FTS5-only (e.g. air-gapped or corporate-proxy network).
• MEMORY_EMBED_MODEL=... — swap in a different Transformers.js feature-extraction model.
• MEMORY_EMBED_CACHE_DIR=... — override the Transformers.js cache location.
• MEMORY_EMBED_DTYPE=fp32|fp16|q8|q4 — model quantization (default q8).

Lifecycle + Reflection (v2.1.0+)

v2.1 closes the gap between "store a fact" and "manage a memory over time". The schema has carried archived, lifecycle_state, valid_from, and valid_to since v1, but no tool exposed them. Now four tools do.

Bi-temporal asOf — "what did I know on date X?"

memory_entity_open({ id: "...", asOf: "2026-04-15" })

Returns the entity plus the observation set whose validity window contained 2026-04-15. The filter is valid_from <= asOf AND (valid_to IS NULL OR valid_to > asOf). Accepts any format SQLite's datetime() recognizes: ISO 8601 (2026-04-15T00:00:00Z), SQLite-style (2026-04-15 00:00:00), or date-only (2026-04-15). Without asOf you get the legacy live-view (every observation with valid_to IS NULL).

Design choice — valid-time only, not full bi-temporal. SQL:2011, XTDB, Datomic offer two-axis bi-temporal (valid-time × transaction-time). We do valid-time only; transaction-time lives passively in created_at but isn't queryable as a separate axis. For a local AI-memory product the question is "what did the AI know about X on date Y" — that's valid-time. Full bi-temporal matters for regulated audit trails (insurance, banking) — if you need it, reach for XTDB.

Scale note. The asOf predicate wraps valid_from in SQLite's datetime() for format-robust comparisons, which means the planner can't use an index on the column directly. For corpora of <1000 observations per entity (typical) the scan is sub-millisecond. If you have an entity with 10k+ observations, add an expression index — CREATE INDEX idx_obs_valid_from_dt ON entity_observations(datetime(valid_from)) — and the predicate becomes sargable again.

Contradiction scanner — LLM-free

memory_contradictions({ minCosine: 0.75, minConfidenceDrift: 0.2 })
memory_contradictions({ entityId: "...", limit: 20 })

Surfaces observation pairs that are semantically very close (cosine similarity above minCosine) but disagree on either:

• negation marker XOR — one side asserts, the other negates (regex covers EN / DE / ES / Catalan).
• confidence drift — same surface claim, very different confidence values.

LLM-free on purpose — the no-API-key promise holds. The heuristic is conservative; the AI client judges. Pure duplicates (no negation, no confidence drift) are not flagged. The cosine math runs in SQL via vec_distance_cosine, so the extension must be loaded; on platforms where it isn't, the tool returns VECTOR_DISABLED with a clear message instead of degrading silently.

Calibration. Default minCosine = 0.75 follows 2026 retriever-tuning literature (SparseCL on Arguana; Milvus threshold-tuning guidance) which finds a sharp false-positive rise below 0.7. Lower to 0.6 for recall-heavy use, raise to 0.85 for precision-heavy. The negation regex covers EN / DE / ES / Catalan / Portuguese / Italian / French — the seven languages multilingual-e5-small handles strongest.

Supersede — retire a stale fact (v2.2)

memory_observation_supersede({ observationId: "..." })
memory_observation_supersede({ observationId: "...", supersededById: "..." })

The scanner finds a stale fact; memory_observation_supersede acts on it. It sets the older observation's valid_to (a tombstone) so the fact stops surfacing in live memory_search and memory_entity_open — but the row stays in the database, so an asOf query still returns it as the belief that was current before the cutoff. This is the Zep fact-supersession pattern: invalidate, never delete. Pass supersededById and the cutoff becomes the newer fact's valid_from (the old fact was true right up until the new one was recorded); pass validTo for an explicit instant, or neither for now. Same-entity and self-supersede guards keep the operation sane, and a second call is idempotent (already_superseded).

Archive + update — lifecycle for learnings

memory_learn_archive({ learningId: "...", reason: "wrong" })
memory_learn_update({ learningId: "...", content: "…", confidence: 0.9 })

archive is a soft delete: the row stays in learnings (with archived = 1, archived_at, and lifecycle_state = 'archived' | 'archived:<reason>'), the embedding stays in vec0 (so asOf-style cross-references can still resolve), but recall / search / the duplicate check all filter it out. Idempotent.

update edits a live (non-archived) learning. If content changes we re-embed in the F4 atomic pattern (compute outside the transaction, write inside one sync db.transaction()). If the embedding write fails or vec is disabled, the now-stale old embedding is purged so cosine search can't surface a vector that no longer represents the live text. Bumps usage_count and sets last_used so an edit counts as a touch.

Trade-off — no audit trail. update overwrites the previous content. The old text is not retained anywhere. This keeps the schema clean; a future version may add memory_learn_history plus an immutable learnings_history table for users who need point-in-time recovery. If you need an audit trail today, memory_learn_archive(reason: "wrong") the old learning and memory_learn the new one as a fresh row — the old text stays in the archived row.

Reflection — what's important right now

memory_reflect({ lookbackDays: 7, staleThresholdDays: 30 })

Aggregation pass over the recent memory stream — Stanford Generative Agents' reflection step, minus the LLM. Returns a structured payload PLUS a Markdown summary covering:

• Most-used learnings — top N by usage_count touched inside the lookback.
• Stale learnings — created longer than staleThresholdDays ago and never recalled. Archive candidates.
• Hot entities — top N by new observations inside the lookback.
• Open decisions — older than the lookback and verified = 0. Follow-up candidates.

The Markdown is for the LLM to read at session start; the structured fields are for downstream automation (Claude Code Hook, n8n workflow) that wants to react without reparsing. Pass project to scope to one project.

Sleeptime via hooks. Letta / Zep / Mem0 run reflection in a background "sleeptime" loop. We run on-demand because we're a stateless stdio daemon — but you get sleeptime semantics for free by wiring a Claude Code SessionStart or SessionEnd hook (or an n8n cron, or a crontab entry) that calls memory_reflect. The summary lands in the LLM's context at the same time as your memory_session_start snapshot. Zero new infrastructure.

Tools (25)

Sessions

memory_session_start -- Call this first in every conversation. Loads context from your last 3 sessions (summaries, recent learnings) so your AI knows what you were working on. Optional project parameter to scope sessions by project.

memory_session_end -- Call at the end to save a summary. Pass a summary string describing what was accomplished. The next session auto-loads this. Without arguments it closes the active session.

Learnings

memory_learn -- The core tool. Stores a piece of knowledge with a category and content. Categories: pattern (recurring success), mistake (what went wrong), insight (strategic realization), research (external knowledge), architecture, infrastructure, tool, workflow, performance, security. The duplicate gatekeeper checks whether the identical content is already stored. If so, it bumps the usage counter instead of creating a duplicate. To extend an existing entry rather than add a new one, use memory_learn_update with its id. Optional: tags, confidence (0-1), project, memoryType (episodic or semantic, auto-classified if omitted).

memory_recall -- Quick search on learnings only. Pass a query string for keyword search, or omit it to get the most recent learnings. Good for "what did I learn about X" questions. Use limit to control how many results come back (default 10).

memory_search -- Unified search across everything: learnings, decisions, entities, and observations. Uses FTS5 with bm25 ranking. Multi-word queries match any of the words and rank by relevance. Use types array to filter (e.g. ["learning", "decision"]). This is the broadest search tool.

memory_learn_archive (v2.1+) -- Soft-delete a learning. The row stays in the DB (so asOf queries that reference it still resolve) but never resurfaces in recall or search. Optional reason is stored on lifecycle_state as archived:<reason>. Idempotent — calling twice returns already_archived.

memory_learn_update (v2.1+) -- Edit a live learning (content / confidence / tags). At least one field is required. Bumps usage_count + last_used so an edit counts as a touch. Re-embeds atomically when content changes (F4 pattern: compute outside the transaction, write inside one sync db.transaction()). Rejects edits to archived learnings with code: 'ARCHIVED'.

memory_learn_bulk (v2.2+) -- Batch-insert up to 500 learnings in one atomic call. Each item takes the same fields as memory_learn. Exact-duplicate-only gatekeeper (no fuzzy merge) so the result is deterministic insert-or-skip; intra-batch duplicate contents collapse onto the first occurrence. Only the new contents are embedded, in one batched model forward pass. For restoring a backup, seeding a fresh DB, or migrating from another system.

When to use recall vs search: Use recall when you want learnings specifically. Use search when you want to find anything across all types, including entities and decisions.

Decisions

memory_decide -- Records a decision with structured context. Parameters: title (what was decided), decision (the choice made), reasoning (why), alternatives (what else was considered). Optional: confidence, project, tags. This is useful for looking back at past decisions months later and understanding why you chose something.

Knowledge Graph

memory_entity_observe -- Record a fact about a person, project, company, tool, or any other entity. If the entity does not exist yet it gets created automatically. Parameters: entityName, entityType (person, project, company, tool, concept, etc.), content (the fact). Observations are bi-temporal, meaning they can be superseded over time without losing history.

memory_entity_search -- Fuzzy search across entity names and their observations. Finds "Claude" even if you search for "claude ai". Optional entityType filter to narrow results.

memory_entity_open -- Load a full entity view: the entity itself, all its current observations, and all its relations to other entities. Search by name or id. v2.1: optional asOf parameter for a bi-temporal point-in-time view — "what did I know about this entity on date X?"

memory_entity_relate -- Create a typed, directed edge between two entities. Parameters: fromEntityId, toEntityId, relationType (e.g. "works_at", "uses", "created", "depends_on"). Optional weight (0-1). Build a graph of how things connect.

memory_contradictions (v2.1+) -- LLM-free scanner that surfaces observation pairs with high cosine similarity but disagreeing on negation markers or confidence. The AI client (Claude / Cursor) judges the candidates. Optional scope: entityId or entityName + entityType. Knobs: minCosine (default 0.75), minConfidenceDrift (default 0.2), limit (default 20). Requires sqlite-vec — returns VECTOR_DISABLED if not loaded.

memory_observation_supersede (v2.2+) -- The execution arm for memory_contradictions: retire a stale observation by setting valid_to (a tombstone). The row stays in the DB so an asOf query still surfaces it, but it drops out of live memory_search and memory_entity_open. Pass supersededById to set the cutoff to the newer fact's valid_from (Zep fact-supersession), validTo for an explicit instant, or neither for now(). Same-entity + self-supersede guards; idempotent.

Recommended entity types: person, project, company, tool, concept, service, team. Use whatever makes sense for your domain.

Reflection

memory_reflect (v2.1+) -- Aggregation pass over the recent memory stream. Returns structured data plus a Markdown summary covering: most-used learnings (top N by usage_count touched in lookback), stale learnings (created > staleThresholdDays ago, never recalled), hot entities (top N by new observations in lookback), open decisions (verified = 0, older than lookback). LLM-free — Stanford Generative Agents' reflection step without the API call. Defaults: lookbackDays: 7, staleThresholdDays: 30, limit: 5. Optional project filter.

memory_insights -- Overview stats: how many days of memory, total sessions, learnings, decisions, entities. Category breakdown and entity type breakdown. Good for "what does Claude know about me" moments. Optional project filter.

memory_profile -- Store personal info locally. Use set to store fields (name, role, preferences, language, timezone), use get to retrieve them. Your AI can read this at session start to personalize its behavior.

memory_guide -- Built-in help. Topics: quickstart (how to get started), session (session workflow), search (how search works), entities (knowledge graph explained), learn (learning categories), privacy (where data lives, what is collected).

Portability (v2.2+)

memory_export -- Dump the whole memory — learnings, decisions, entities, observations, relations, sessions, profile, goal — to a versioned, camelCase JSON envelope (format: "studiomeyer-memory-export", version: 1). Embeddings are not exported; they're re-derived on import, so the file stays small and model-agnostic. Flags: includeSessions (default true), includeArchived (default true). You own your data — it's already a single SQLite file, and now it's a portable document too.

memory_import -- Ingest a memory_export envelope ({ data: <envelope> }). Purely additive and idempotent: every write is INSERT OR IGNORE on the source id, so re-importing the same file is a no-op and it never clobbers an existing row. FK-safe ordering with dangling-reference skips (a counted skip, never a throw). Re-embeds on the fly. There is no replace mode by design — to wipe a local store, delete memory.sqlite. The same envelope also imports into the hosted StudioMeyer Memory tier, so this is your on-ramp when you outgrow a single machine.

Tips

• Start with sessions and learnings. Just calling memory_session_start at the beginning and memory_learn when something important comes up already gives you 80% of the value.
• Use entities for people and projects. When you mention a colleague, client, or project repeatedly, create an entity. Over time you build a knowledge graph that your AI can traverse.
• Decisions are underrated. Three months from now you will not remember why you chose Postgres over SQLite for that project. memory_decide captures the reasoning.
• Let your AI drive. Once the tools are available, your AI will naturally start using them. You do not need to call tools manually. Say "remember this" and it calls memory_learn. Say "what do you know about Sarah" and it calls memory_entity_search.
• Back up your SQLite file. It is a single file. Copy it to a USB drive, Dropbox, wherever. You can also open it with any SQLite browser to inspect what your AI has learned.

Features

• Knowledge Graph -- not just flat text. Entities, bi-temporal observations, typed relations.
• Duplicate Guard -- FTS5 similarity check prevents storing the same thing twice. Usage counter instead.
• Session Context -- auto-loads last 3 sessions on start. Your AI picks up where you left off.
• Decision Tracking -- log decisions with reasoning and alternatives. Unique among memory servers.
• Full-Text Search -- FTS5 with bm25 ranking across learnings, decisions, entities, observations.
• Single SQLite File -- one file, portable, backupable, deletable. WAL mode for concurrent access.
• Zero Config -- npx and done. No Docker, no Postgres, no Redis, no API keys.

Where your data lives

Everything in one SQLite file. Back it up, move it, delete it -- it's yours.

Override: MEMORY_DB_PATH=/your/preferred/path.sqlite

Privacy

• Your data never leaves your machine
• No telemetry, no phone-home, no analytics
• No account required, no API keys needed
• Open source -- read every line of code

Comparison

Where we stand out: the only local, MIT-licensed, API-key-free memory MCP shipping hybrid retrieval (BM25 + vector cosine via RRF) with multilingual embeddings and one-click installers for every desktop OS. Decision tracking remains unique to us.

local-memory-mcp vs. StudioMeyer Memory

Two products, same team, different use cases:

Start local. Upgrade when you need teams, multi-device sync, or cross-encoder rerank — memory_export produces an envelope that imports straight into the hosted tier, so you carry your whole memory across with you.

Also by StudioMeyer

Security

See SECURITY.md for the threat model, reporting process, and notes on known SAST scanner false positives. In particular: db.exec(schema) in src/db/client.ts is better-sqlite3's SQL-string executor, not child_process.exec — some pattern-based scanners flag it without import resolution. The repo contains zero shell-execution code (verify with grep -rn child_process src/).

Contributing

Issues and PRs welcome. See CONTRIBUTING.md.

About StudioMeyer

StudioMeyer is an AI and design studio based in Palma de Mallorca, working with clients worldwide. We build custom websites and AI infrastructure for small and medium businesses. Production stack on Claude Agent SDK, MCP and n8n, with Sentry, Langfuse and LangGraph for observability and an in-house guard layer.

License

MIT

Built by StudioMeyer -- AI-first web studio from Mallorca.