Seahorse

Persistent, bi-temporal memory for LLM agents — local-first, MCP-native, Obsidian-readable.

pip install seahorse-memory
seahorse init myvault && seahorse remember "Sergio lives in Madrid"
seahorse recall "where does Sergio live?"

Why

LLM agents start every session from zero. The context window is not memory: it is a scratchpad that resets, and it is too small to hold what an agent has learned across weeks of work. The tools that try to fix this have their own problems:

• They forget badly. Most memory systems accumulate facts forever and never resolve contradictions — an agent "remembers" that a user lives in Madrid and Barcelona at the same time, with no way to know which is current.
• They are opaque. Memory lives in a proprietary database the human cannot read, edit, or audit. If the agent is wrong, there is no way to correct it.
• They are expensive to feed. Every episode goes through an LLM, so writing thousands of small facts costs real money.
• They lock you in. Adopting a memory system often means adopting its runtime, its provider, or its ecosystem.
• Their benchmarks are not trustworthy. The field's own numbers are hard to reproduce: the LOCOMO benchmark has 6.4% wrong gold answers, Mem0's reproduction is broken (issue #2800), and MTEB embedding scores do not predict memory-retrieval performance (LMEB, arXiv 2603.12572).

Seahorse is a different approach: an open, portable, bi-temporal memory standard that an agent writes to and reads from, that a human can read and correct, and that does not lock you into any runtime or provider.

Who it's for

• Developers building agents (Claude Code, Cursor, Codex, or your own) who want the agent to remember decisions and context across sessions.
• Obsidian power users who want their notes to be more than a static archive — a knowledge base an agent can query and maintain.
• Teams that want portable memory — a format they can migrate between vendors without replaying history.

Use case: Claude Code with persistent memory

The fastest way to see Seahorse is to give Claude Code a memory that survives between sessions. Three steps:

1. Capture sessions. seahorse setup installs the observer hooks into ~/.claude/settings.json; seahorse observe start runs the capture worker. Every session is recorded as episodes — skip-first (near-zero cost), redacted, with a deterministic summary.

seahorse setup
seahorse observe start

2. Recall across sessions. The SessionStart hook injects seahorse context into the next session, so the agent starts with what it learned before. Ask directly with seahorse recall:

seahorse context
seahorse recall "what did we decide about the API design?"

3. Bring your existing memory. If you already use claude-mem, seahorse import migrates its observations into canonical episodes — no replay, no lock-in:

seahorse import --mode commit

The key difference: the agent writes into the same vault you edit in Obsidian. Every episode is a markdown file with YAML frontmatter — readable, editable, diffable in git, and auditable by a human. The agent's memory is not a black box; it is your notes.

How it works

graph LR
    A[Claude Code / any MCP agent] -- stdio MCP io.seahorse.memory/v1 --> S[seahorse-mcp]
    S --> E[Bi-temporal engine]
    E --> DB[(sqlite3 + sqlite-vec + FTS5)]
    E --> V[Obsidian vault: markdown + F3.1 frontmatter]
    H[Human in Obsidian] --> V

An agent talks to seahorse-mcp over stdio MCP. The engine stores every episode twice: once in a single-file SQLite database (sqlite-vec for vector search, FTS5 for full-text), and once as a markdown file with F3.1 frontmatter in the vault. The human edits the same markdown. The format is versioned and documented in docs/f3.1-format.md.

Quickstart

# Install (PyPI):
pip install seahorse-memory
# …or with uv:
uv tool install seahorse-memory
# For hybrid semantic retrieval (FastEmbed ONNX, downloads mE5-small on first
# embed): pip install "seahorse-memory[embeddings]"
# For the multi-LLM extraction path (LiteLLM): pip install "seahorse-memory[llm]"

# Create a vault and write your first episode:
seahorse init myvault
seahorse remember "Sergio lives in Madrid" --title home
seahorse recall "madrid"

# Improve and forget (append-only; history is preserved):
seahorse improve <ep_id> "Sergio lives in Barcelona" --reason correction
seahorse forget <ep_id> --reason done

# Session capture, context, and consolidation:
# Install the observer (writes [observe] + merges the Claude Code hooks into
# ~/.claude/settings.json):
seahorse setup
# Start the observer (unix socket + worker), then the next session is captured
# automatically (skip-first, redacted, deterministic summary):
seahorse observe start
seahorse observe status
# Bootstrap context by recency (the SessionStart hook injects this):
seahorse context
# Distill recurrent episodes into semantic knowledge notes (N≥3, idempotent):
seahorse consolidate
# Remove the observer:
seahorse setup --uninstall

# Serve an agent over stdio MCP (io.seahorse.memory/v1):
seahorse-mcp --vault myvault
# …equivalently:
seahorse mcp --vault myvault

The seahorse console script is for humans and shell scripts; seahorse-mcp is for agents. The seahorse mcp subcommand invokes the same stdio server as seahorse-mcp, so both agent entry points are equivalent.

Prerequisites

• Python ≥ 3.11 (any recent 3.11/3.12/3.13 works). The interpreter's sqlite3 must support enable_load_extension (sqlite-vec needs it); most standard builds do — seahorse doctor reports it as a FAIL if not.
• Obsidian is optional. Seahorse runs on any directory of markdown — seahorse init creates a .seahorse/ sidecar in a plain folder. Obsidian is a human-facing editor for the same folder; its .obsidian/ directory is ignored by Seahorse, never required.

Migrating a legacy Obsidian vault

A vault of pre-existing Obsidian notes (no frontmatter, or legacy tags/ created frontmatter) is not yet in the canonical format — seahorse index rebuild fails honestly on those notes. seahorse frontmatter migrate converts them:

# Preview: classify every note, write nothing (always exit 0):
seahorse frontmatter migrate --vault myvault --dry-run
# Apply: convert legacy notes, leave canonical notes untouched, refuse
# incompatible notes:
seahorse frontmatter migrate --vault myvault
# Rebuild the sidecar index from the converted notes:
seahorse index rebuild --vault myvault

Apply exits 97 when incompatible notes block a full migration — the manifest summary is printed first so the operator sees which notes need manual resolution. --resume skips notes unchanged since the last manifest; --batch-size sets the manifest checkpoint cadence. Migration works before seahorse init (it only touches .md files + the manifest).

First run: the semantic-embedding model (mE5-small, ~235MB) downloads lazily on the first remember/recall — the CLI announces it so the first call doesn't look hung. seahorse status shows the active retrieval regime (hybrid RRF (model cached) vs current-state listing — install seahorse-memory[embeddings] for semantic recall).

Compared to other memory tools

A comparison of verified facts, not a ranking. Sources: the project's state-of-the-art analysis (see the research notes and the claims cited below).

	Seahorse	mem0	Letta / MemGPT	Zep / Graphiti	claude-mem	LangMem
Portable open format	✓ F3.1 spec	✗ proprietary	✗ runtime-bound	✗	✗ own schema	✗
Human-readable layer	✓ Obsidian vault	✗	✗	✗	✗	✗
Bi-temporal (point-in-time)	✓	~	~	✓ Graphiti	✗	✗
Local-first, zero-infra	✓	~	~	✗ cloud-only	✓	~
Reproducible benchmark	✓ harness in-repo	✗ #2800	—	—	—	—
License	Apache-2.0	Apache-2.0 (open-core)	Apache-2.0	Apache-2.0	AGPL	Apache-2.0

Legend: ✓ yes · ~ partial · ✗ no · — not verified.

The two facts that matter most: mem0 paywalls the features that produce its benchmark numbers, and Zep abandoned self-host for cloud-only. Seahorse is local-first by default, publishes its benchmark harness in the repo, and keeps the memory format portable so you are never locked in.

Benchmark

Seahorse ships a reproducible benchmark harness (LMEB-S, a subsample of the LongMemEval benchmark) and publishes its own numbers — with caveats. The point is not a leaderboard; it is an honest, reproducible measurement.

Metric	Value	Note
recall@10	0.13	knowledge-update slice: 0.44
ndcg@10	0.11
mrr	0.13	knowledge-update slice: 0.47
precision@10	0.02
token efficiency	0.998	51.5M tokens full-context → 121K measured
latency p95 (INDEX)	42 ms	retrieval-only, no rerank

Caveats: the run uses a subsample (n≈470–500 questions, not the full dataset); relevance is judged by a small LLM without human validation; and it measures retrieval only, not the agent's final answer. A cross-encoder rerank was tested and rejected — it degraded recall@10 to 0.11 with 1.2s latency. Full methodology and reproduction commands in docs/benchmark.md.

FAQ

What is an episode? A single memory record: a markdown file with YAML frontmatter carrying two time axes (valid_at — when it became true, and created_at — when it was recorded), provenance, and a cognitive type. The format is versioned and documented in docs/f3.1-format.md.

Why Obsidian? Because the human is part of the memory system. The agent writes into the same vault you edit — markdown is readable, diffable in git, and auditable. If the agent is wrong, you correct the note, not a database.

How is this different from claude-mem? claude-mem stores session observations in its own schema. Seahorse is an open, bi-temporal standard with a portable format and a human-readable layer — and seahorse import migrates claude-mem observations into canonical episodes, so it is a bridge, not a competitor.

Do I need an LLM? No. The deterministic skip-path is the default for the bulk of writes (near-zero cost). LLM extraction is optional (seahorse-memory[llm]) and reserved for the few episodes that justify it.

Is it free? Yes. Apache-2.0, local-first, zero-infra. A managed SaaS and enterprise tier are planned for the future (see the project's strategy notes).

How do I contribute? See CONTRIBUTING.md for the development setup, test/lint commands, and pull request workflow.

Roadmap

See ROADMAP.md for what is built, what is next, and the direction of the project. Release history lives in CHANGELOG.md.

The agent surface — 7 memory-native primitives + 5 procedural/read-only tools

Exposed over stdio MCP (io.seahorse.memory/v1, protocol pinned 2025-11-25) and mirrored on the CLI. These are memory primitives, not generic CRUD: an agent calls remember / recall / improve / forget the way a human would talk about memory.

The 7 primitives (write + retrieve):

Primitive	What it does
remember	Record an episode (body, source, optional title/subject).
recall	INDEX level — the current-state listing, clamped to top_k.
recall_timeline	TIMELINE level — the supersedes chain around an anchor episode.
recall_full	FULL level — the hydrated episode with all provenance.
improve	Supersede an episode with a corrected one (append-only).
forget	Soft-delete an episode (append-only; history preserved).
build_pit	Build a point-in-time projection (all-None → current state).

Plus 5 procedural / read-only tools (skills + facade introspection):

Tool	What it does
skill_add	Create a procedural skill (deterministic, near-zero cost).
skill_show	Show a skill's gated body (trust gate).
freshness_view	Freshness snapshot of an episode (age, stale, pending_ingest).
audit_log	Audit events for an episode (write-path history).
follow_supersedes_chain	The supersedes closure for an episode (version history).

Three retrieval levels give progressive disclosure: a cheap listing first (INDEX), the chain on demand (TIMELINE), and the full record only when needed (FULL). This keeps the common path cheap.

What works

• Bi-temporal, append-only episode store on stdlib sqlite3 + sqlite-vec (FTS5
  • vec0). Auto-migrating schema.
• The 7 memory-native primitives plus 5 procedural / read-only tools, on both the CLI and stdio MCP (12 tools total).
• Progressive disclosure (INDEX / TIMELINE / FULL) and point-in-time projection.
• Hybrid semantic retrieval: recall ranks by relevance — sqlite-vec kNN + FTS5 BM25 fused with Reciprocal Rank Fusion, with point-in-time routing (state_at / known_at) when a real embedder is wired. The write path and seahorse index rebuild populate vec0/FTS (best-effort — an embedder failure never fails the episode write).
• Honest degrade: without the embeddings extra (or with no vectors populated), recall falls back to the current-state listing (score 0.0, no ranking) and point-in-time recall is refused — the engine keeps working without ranking.
• LLM extraction: a real multi-LLM path (ollama / gemini / groq / openrouter / openai / anthropic / deepseek / vllm, local-first) with a strict schema validator + repair loop, retry/fallback chain, and an operative cost cap (local and free-tier models price at $0). seahorse init --llm bootstraps it; the skip-path stays the near-zero-cost default for the bulk of writes.
• Local-first CI gate: the real extraction path runs in CI against the weakest model of the family (ollama/qwen3:0.6b) so the validator + repair must carry the load — the path does not silently depend on native structured outputs or a strong model.
• Supersession (improve) and soft-delete (forget) with full history preserved.
• Frontmatter import/export for the Obsidian vault layer (markdown as the human-readable, portable on-disk contract).
• Legacy-vault migration: seahorse frontmatter migrate converts legacy Obsidian notes with a --dry-run preview, --resume, and honest exit 97 when incompatible notes block a full migration.
• Honest exit codes and a structured {"error": {...}} envelope on stderr, so agents and scripts can branch on seahorse_code / cli_code deterministically.

A few CLI commands are wired but intentionally return exit 75 with a reason (expire, revalidate, index verify), so the surface is honest about what is not implemented yet rather than silently no-op'ing. Batch distillation (consolidated) is schema-valid but not built; llm_partial stays fully reserved.

Stack

• Python ≥ 3.11. stdlib sqlite3 + sqlite-vec for storage (zero-infra single file; the vec0 virtual table + FTS5).
• numpy for the embedding blob shape.
• Pydantic v2 for the canonical Episode contract (core type system).
• Typer for the CLI surface (humans and scripts). Confined to seahorse.cli.
• stdio JSON-RPC 2.0 for the MCP agent surface (hand-rolled framing, stdlib-only seahorse.mcp package — import seahorse.mcp does not load Typer).
• ruamel.yaml + python-frontmatter, confined to the frontmatter adapter.
• FastEmbed ONNX + onnxruntime (embeddings extra, NOT in the default install): the mE5-small bundle defaults to model_O4.onnx (fp32, ~235MB) — no int8/fp16 artifact is portable to Apple Silicon, and an open standard must run on Windows/Linux/macOS. A portable int8 bundle is a measured follow-up.
• LiteLLM (llm extra, NOT in the default install): unifies the 100+ provider surface for the LLM extraction path. Without the extra, seahorse.llm still imports (contract + StubLLMClient) and the real path degrades llm→skip with a setup hint.

The FastAPI / SQLAlchemy / Postgres stack is planned for a later multi-agent tier (Postgres + pgvector). The README states what ships now, not the target architecture.

Testing

• Unit + integration: uv run pytest (coverage ≥ 80% gate).
• Fresh-user e2e: scripts/e2e-fresh-user.sh — the full install → init → core CLI → embeddings → LLM → import → MCP flow from a clean, isolated HOME (never touches the real ~/.claude / ~/.claude-mem).
• Environment matrix: scripts/e2e-matrix.sh — the fresh-user flow across environment combinations (install method × extras × Obsidian × Ollama × online/offline × vault state × concurrency). --ci-subset runs the CI-safe combos (core_min + uv_sync_dev); --list shows all combos.
• Core stress: scripts/stress-core.sh — ingest 1000+ episodes, recall --top-k 100 p95 ≤ 250ms (in-process INDEX budget), concurrent single-writer, reindex, idempotent import, improve/forget chain.

Contributing

Contributions are welcome. See CONTRIBUTING.md for the development setup, test/lint commands, and pull request workflow. Release history lives in CHANGELOG.md.

License

Apache-2.0. See LICENSE.

Current status

v0.6.0. The memory engine works end-to-end from a clean install: write episodes, recall them with hybrid semantic retrieval, extract with a real multi-LLM path (local-first, CI-gated), improve and forget them, and serve an agent over stdio MCP. Recall ranks by relevance when vectors are populated and the embedder is wired, and honestly degrades to a current-state listing otherwise. seahorse import migrates claude-mem observations to episodes, and an opt-in recency ranking signal is available. See What works and ROADMAP.md for what is next.