oneMEM

One memory. Every AI. You own it.

Local, structured memory for AI agents — in one SQLite file on your machine.

oneMEM gives AI tools a shared local memory. It distills useful context into compact atomic facts and surfaces the minimum memory sufficient for a query. Every connected agent reads and writes the same SQLite file.

AI agents ───┐
Code editors ┼── MCP ── oneMEM ── ~/.onemem/onemem.db
Local tools ─┘

Why oneMEM?

Quick Start

Requires Python 3.11+ and an API key for any LLM provider. Embeddings run locally.

# Install
uv tool install "onemem[all]"

# Setup (walks you through provider, key, model, capture, MCP wiring)
onemem init

# Try it
onemem add "Chose SQLite because it needs zero operations and one-file backups."
onemem ask "What storage did I choose, and why?"

Architecture

graph TB
    subgraph INPUTS ["Inputs"]
        CLI["CLI<br/>onemem &lt;command&gt;"]
        MCP["MCP Server<br/>onemem-mcp"]
        API["HTTP API<br/>FastAPI /events"]
        WATCH["Watch<br/>Claude Code / Codex transcripts"]
    end

    subgraph WRITE ["Write Path"]
        INTAKE["1. ingest_event()<br/>chunk - dedup by content hash - store"]
        EXTRACT["2. extract_entities()<br/>LLM reads event - atomic facts + named entities"]
        RECONCILE["3. reconcile + store<br/>normalize entities - link fact_entity_edges - store facts"]
        EMBED_W["4. embed_facts()<br/>bge-base-en-v1.5 768-d local embedding"]
    end

    subgraph SQLITE ["SQLite - ~/.onemem/onemem.db"]
        EVENTS[("events<br/>raw content, append-only")]
        EXTR[("extractions<br/>provenance ledger")]
        FACTS[("facts<br/>atomic claims")]
        ENTITIES[("entities<br/>canonical names")]
        EDGES[("fact_entity_edges<br/>which entities each fact mentions")]
        EMBED[("fact_embeddings<br/>sqlite-vec vec0, cosine")]
        FTS[("facts_fts<br/>FTS5 keyword index")]
    end

    subgraph READ ["Read Path"]
        PARAMS["1. LLM param extraction<br/>question - topic keywords + date range"]
        RETRIEVE["2. Deterministic Retrieval"]
        VECTOR["Vector Door<br/>cosine similarity"]
        KEYWORD["Keyword Door<br/>FTS5 BM25"]
        ENTITY_D["Entity Door<br/>fact_entity_edges"]
        FUSION["Fusion<br/>magnitude noisy-OR"]
        CUT["3. Adaptive Cut<br/>score-curve ratio, bounded 10 to limit"]
        COLLAPSE["Source Collapse<br/>if facts - raw event tokens - return raw"]
        SYNTH["4. LLM Synthesis<br/>optional natural-language answer"]
    end

    CLI --> INTAKE
    MCP --> INTAKE
    API --> INTAKE
    WATCH --> INTAKE

    INTAKE --> EVENTS
    EVENTS --> EXTRACT
    EXTRACT --> FACTS
    EXTRACT --> ENTITIES
    EXTRACT --> EXTR
    RECONCILE --> EDGES
    EMBED_W --> EMBED
    FACTS --> FTS

    CLI --> PARAMS
    MCP --> PARAMS

    PARAMS --> RETRIEVE
    RETRIEVE --> VECTOR
    RETRIEVE --> KEYWORD
    RETRIEVE --> ENTITY_D
    VECTOR --> FUSION
    KEYWORD --> FUSION
    ENTITY_D --> FUSION
    FUSION --> CUT
    CUT --> COLLAPSE
    COLLAPSE --> SYNTH

    EMBED --> VECTOR
    FTS --> KEYWORD
    EDGES --> ENTITY_D

Key design principles:

• Append-only — raw events are never overwritten; facts are only ever added
• Deterministic retrieval — no LLM in the read path; same query always returns the same result
• Small models at the edges — LLM only at write time (distill) and optionally at read time (synthesize)

User Flow

%%{ init: { 'theme': 'dark', 'themeVariables': { 'actorBkg': '#7C3AED', 'actorTextColor': '#fff', 'actorBorder': '#9F67FF', 'signalColor': '#E2E8F0', 'signalTextColor': '#E2E8F0', 'noteBkgColor': '#1E293B', 'noteTextColor': '#E2E8F0', 'noteBorderColor': '#475569', 'rectBkgColor': '#0F172A', 'rectBorderColor': '#334155', 'rectTextColor': '#E2E8F0', 'sequenceNumberColor': '#fff' } }%%
sequenceDiagram
    actor User
    participant CLI as CLI / MCP
    participant Core as oneMEM Core
    participant LLM as LLM Provider
    participant DB as SQLite

    rect rgb(15, 23, 42)
        Note over User, DB: Write — ingest and distill
        User ->> CLI: onemem add "note"
        CLI ->> Core: ingest_event()
        Core ->> DB: store raw event
        Core ->> LLM: extract facts + entities
        LLM -->> Core: ExtractionResult
        Core ->> DB: store facts, entities, edges
        Core ->> Core: embed facts (768-d, local)
        Core -->> CLI: event_ids
    end

    rect rgb(15, 23, 42)
        Note over User, DB: Read — retrieve and answer
        User ->> CLI: onemem ask "question?"
        CLI ->> LLM: extract search params
        LLM -->> CLI: {text, start, end}
        CLI ->> Core: retrieve(text, start, end)
        Core ->> DB: vector + keyword + entity search
        DB -->> Core: matched facts
        Core ->> Core: fusion then adaptive cut
        Core -->> CLI: facts with scores
        CLI ->> LLM: synthesize answer from facts
        LLM -->> CLI: AskAnswer
        CLI -->> User: natural language answer
    end

    rect rgb(15, 23, 42)
        Note over User, DB: MCP — agent background write
        User ->> CLI: AI agent conversation
        CLI ->> Core: onemem_log(content)
        Core ->> DB: store raw event
        Note right of Core: background processor<br/>extracts facts later
    end

Command Flow

Commands

MCP Setup

oneMEM works with any MCP client that supports local stdio servers.

# Claude Code
claude mcp add --scope user onemem -- "$(command -v onemem-mcp)"

# Codex
codex mcp add onemem -- "$(command -v onemem-mcp)"

onemem init automatically detects and wires Claude Code and Codex during setup.

MCP Tools

Supported Providers

Embeddings always use bge-base-en-v1.5 (768-d) running locally — no API key needed.

Benchmarks

Measured on a 100-instance stratified sample of LongMemEval-S:

Configuration

Edit ~/.onemem/config.toml (or use onemem config set):

[model]
provider = "openrouter"
model = "google/gemini-3.5-flash-lite"

[spend]
max_run_cost_usd = 20.0    # hard ceiling per batch import

[retrieval]
default_limit = 30         # max facts returned per recall
neighbour_max = 20         # neighbour facts gathered around a match

[ingestion]
concurrency = 20           # parallel LLM workers during bulk import

Where data lives

Development

git clone https://github.com/shashank-tomar0/onemem.git
cd onemem
uv sync --all-extras
uv run pytest -q                    # 144 passing
./scripts/dev-onemem doctor         # run with isolated dev home

Project Structure

onemem/
├── cli/                  # Click CLI (init, add, ask, watch, ...)
├── api/                  # FastAPI HTTP API
├── providers/            # LLM + embedding implementations
│   ├── openai_compat.py      # OpenAI-compatible endpoints
│   ├── anthropic.py          # Anthropic native API
│   └── local_embedding.py    # bge-base-en-v1.5
├── mcp_server.py         # MCP server (onemem_recall + onemem_log)
├── fact_retrieval.py     # Deterministic hybrid search
├── pipeline.py           # Ingest + process orchestration
├── entity_extractor.py   # LLM-based entity + fact extraction
├── schema.sql            # SQLite schema
└── config.py             # All tunable settings

License

MIT — Based on Meniscus by magic_bubblez.