Genome

GENOME

Open memory for AI agents. Same answer accuracy as Mem0 - but ~1,000× cheaper to store, runs fully offline, and keeps an auditable record.

Papers: Do Agents Need an LLM to Remember? (the core evaluation, 2026) and What Does Each Memory Feature Buy? (a measured audit of all five optional features, wins and failures alike, 2026). PDFs in papers/; result tables in benchmarks/AUDIT-RESULTS.md.

Most agent-memory tools (like Mem0) call an LLM on every message to decide what to remember. That's the slow, expensive part - and GENOME's bet is that you don't need it. GENOME just embeds each message locally: no LLM, no API, no network in the write path.

Benchmarked honestly on public datasets (LoCoMo, LongMemEval), GENOME answers just as accurately as Mem0 - while storing memories for a tiny fraction of the cost and running completely offline.

Honest up front: on answer accuracy, GENOME *ties* Mem0 - we do not claim to beat

it there (six independent benchmark configurations confirm parity, none significant in

either direction). The advantage is cost, speed, offline operation, and a

temporal/auditable record Mem0 can't produce.

See it work

Every frame is real output from examples/demo_timeline.py, captured by tools/render_demo_gif.py. Run it yourself, no API key required:

python examples/demo_timeline.py

The interesting part is step 3. The same question gets three different correct answers depending on *when* you ask about, because the store keeps when each fact became true rather than overwriting it:

The "thinking about maybe moving to Denver, nothing decided" turn is stored but never becomes an answer: it is a plan, not a durable fact.

How it works

The write path is deliberately dumb and cheap. All the intelligence happens at read time, when there is a query to focus it.

flowchart LR

M["incoming message"] --> E["local embedder all-MiniLM-L6-v2"] E --> S[("local store SQLite or Postgres")] M -. "optional, opt-in" .-> B["belief extraction (the only LLM call)"] B --> K[("bi-temporal fact log")]

Q["query"] --> R["exact cosine search over this tenant's rows"] S --> R R --> RR["optional cross-encoder rerank"] RR --> A["context for the agent"] Q --> PIT["as-of resolution facts_valid_at(entity, T)"] K --> PIT PIT --> A

style E fill:#0A84FF,color:#fff style S fill:#1c2530,color:#fff style K fill:#1c2530,color:#fff style B fill:#3a3a3a,color:#fff

Write: embed locally, store. About 10 ms, zero LLM calls, zero network calls. The embedding is deterministic -- the same text always yields the same vector, with no sampled extraction step deciding what matters -- so what gets stored is a function of the input, and replaying a journal reproduces that store exactly. (Ids and timestamps are stamped per write, so two independent ingests of the same conversation agree on content and vectors, not on record ids.)

Read: exact cosine search within the tenant's scope (no ANN index to build or update), with an optional local cross-encoder reranker.

Bi-temporal layer (opt-in): records each fact at its domain time, the moment it became true in the world, not the moment it was ingested. That is what makes point-in-time questions answerable even when facts arrive out of order.

Why the record can be re-derived

flowchart TB

subgraph LLM["LLM-extraction memory"] A1["message"] --> A2["LLM decides what matters (sampled, non-deterministic)"] A2 --> A3[("store")] A3 --> A4["replaying the same input can produce a different store"] end subgraph GEN["GENOME"] B1["message"] --> B2["local embedding (deterministic)"] B2 --> B3[("store")] B3 --> B4["replaying the same input reproduces the same store"] end style A4 fill:#5c1f1f,color:#fff style B4 fill:#1f4d33,color:#fff

A record that cannot be re-derived is difficult to audit. That property, not accuracy, is the actual argument for this design.

Don't believe it? Prove it yourself

The cost, speed, and offline claims need no API key - measure them on *your* machine in 60 seconds:

git clone https://github.com/NORTHTEKDevs/genome && cd genome

pip install -e . && python -m genome.verify

The first run downloads the local embedding model (~90 MB, one time) before printing anything, so expect 30-120 seconds of apparent silence on a cold machine. Every run after that is instant.

It writes memories with your outbound network physically blocked and prints a live pass/fail receipt - 0 network calls, 0 LLM calls, single-digit-ms writes, retrieval that works:

  [PASS] Air-gapped write path: wrote 200 memories with every outbound socket blocked -> 0 network attempts, 0 LLM calls

[PASS] Write latency: 7.1 ms/message (Mem0's measured write path: ~2,055 ms + 1 LLM call/message) [PASS] Retrieval works: top hit score 0.598

That receipt covers the cost/speed/offline story only. The accuracy-parity with Mem0 claim is a separate, larger check that needs an LLM key - reproduce it head-to-head on the same questions with your own key via python benchmarks/head_to_head.py (one OpenRouter key works; see benchmarks/RESULTS.md for the n=90 / n=205 runs, the paired significance tests, and the published nulls). The full test suite runs in public CI (badge above). The pitch isn't "trust me" - it's "run it."

Add persistent memory to your agent in one line (MCP)

GENOME ships a fully-local MCP server - cross-session memory for Claude Desktop, Claude Code, or Cursor with no API key and no data leaving your machine:

pip install "genome-memory[mcp]"

{ "mcpServers": { "genome": { "command": "genome-mcp" } } }

Or zero-install via uv: { "command": "uvx", "args": ["--from", "genome-memory[mcp]", "genome-mcp"] }

Tools the agent gets: `remember`, `recall`, `forget`, `reset_memories`. Memories persist locally in ~/.genome/memories.db. Full MCP details ↓

GENOME vs Mem0 at a glance

Every number is measured within one harness - same responder, judge, embedder, and top-k; only the memory layer changes - with paired significance tests. Full detail and per-number provenance: benchmarks/RESULTS.md. Formatted report: benchmarks/GENOME-LoCoMo-Report.pdf.

Why it's ~1,000× cheaper: it never calls an LLM to remember

Storing one message costs one LLM call in Mem0, zero in GENOME (just a local embedding). That's not a benchmark you can argue with - it's arithmetic, and it holds no matter which LLM you price it against. At 10,000 users × 50 messages/day (15M messages/month):

The gap survives the cheapest model and *grows* in production (Mem0 re-sends stored memories to the LLM as the store fills). Reproduce: python benchmarks/tco_project.py (no API key).

It runs air-gapped

GENOME's default embedder is local. We proved the write path is genuinely offline by blocking all network during writes - they still succeed:

• ~10 ms/message, 0 network calls, 0 LLM calls (python benchmarks/local_writepath.py)
• Mem0 can't do this - it needs an LLM API call to ingest.

That makes GENOME usable on-prem, in regulated environments, or fully offline. It's a yes/no capability, not a price point.

How it works

• Write: embed the message locally and store it. No LLM, no network. (~10 ms)
• Read: vector search over your memories, with an optional local cross-encoder reranker for harder queries.
• Optional bi-temporal layer: track how facts change over time and answer "what was true at time T" - see below.

What determinism buys you

Because nothing on the write path interprets your content, GENOME can do things an LLM-ingest memory system cannot do in principle:

• Memory firewall (genome.firewall): tag every write with where it came from (user, agent, tool, web), quarantine low-trust origins from recall, and enforce origin-bound authority - web content can never UPDATE or DELETE what your user said, even when a prompt-injected conflict resolver asks for it. There is also no extraction step for injected content to attack: the write path has no LLM.

  from genome import Memory

from genome.firewall import TrustPolicy

m = Memory(trust_policy=TrustPolicy(recall_min_trust=1)) m.add("I live in Anchorage", user_id="u1", provenance="user") m.add(scraped_page_text, user_id="u1", provenance="web") # quarantined

• Explainable recall (genome.explain): explain_search() reports every candidate's dense score, BM25 rank, fused score, and - when it was not returned - the exact reason (parent-filtered, quarantined, beyond the limit). Two runs agree, so a recall bug can be committed as a regression test instead of a shrug.

• Journal + replay (genome.journal): record every mutation and provably reproduce the store - verify_journal() replays the history and compares canonical hashes. Replay a prefix to roll back; replay into different storage to branch a memory for a what-if run. The journal sits after extraction, so replay is deterministic even if you configured an LLM extractor. Each line chains to its predecessor, so a removed or edited line is detected even when the change cancels out in the final state.

  # Tamper-EVIDENT by default. Pass a key (kept outside the journal's directory)

to make it tamper-PROOF: an unkeyed chain can be recomputed by anyone with

write access, an HMAC chain cannot.

m = Memory(journal="mem.journal", journal_key=os.environb[b"GENOME_JOURNAL_KEY"])

• Multi-agent belief attribution (record_fact(..., believed_by="agent-a")): agents sharing a store keep their own belief timelines - agent B disagreeing does not clobber agent A's fact - and