LLM agents start every session from zero: the context window is a scratchpad that resets. Seahorse gives agents memory that survives across sessions — and gives humans a memory they can actually read and correct.

It is an open, portable, bi-temporal memory standard, not a proprietary store. Every memory ("episode") carries two time axes — when it became true (valid_at) and when it was recorded (created_at) — so contradictions resolve instead of accumulating: an agent never "remembers" you living in two cities at once. Onboarding is one command: seahorse setup configures the vault, the database, Claude Code capture hooks, a self-healing observer, the MCP registration, and agent instructions — and it never fails the caller.

Key Features

• One-command onboarding — seahorse setup configures the whole stack and exits 0 always; every step degrades to a WARN line with the exact fix.
• Not a black box — every episode is a Markdown file with versioned YAML frontmatter (the F3.1 spec) in a vault you can open in Obsidian. The agent writes into the same notes you read and edit.
• Bi-temporal with honest corrections — improve supersedes, forget soft-deletes; full history is preserved and point-in-time views are native.
• Near-zero cost by default — the bulk of writes are deterministic (no LLM call); semantic retrieval runs locally (SQLite + sqlite-vec + FTS5); even distillation uses the agent's own LLM, so no API key is required.
• No lock-in — local-first, zero-infra, Apache-2.0, portable format, and a reproducible benchmark harness published in the repo (with caveats, not leaderboard numbers).

Example Usage

Install (semantic retrieval + multi-LLM extraction extras):

uv tool install seahorse-memory --with "seahorse-memory[embeddings,llm]"

One command, everything configured:

seahorse setup

The agent now remembers across sessions:

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

How It Works

Agents talk to a stdio MCP server (io.seahorse.memory/v1) exposing 14 memory-native tools — remember, recall, recall_timeline, recall_full, improve, forget, and more. The engine stores each episode twice: in a single-file SQLite database (sqlite-vec kNN + FTS5 BM25, fused with Reciprocal Rank Fusion) and as a Markdown file with F3.1 frontmatter in the vault. Humans edit the same Markdown; seahorse doctor --fix diagnoses and repairs the whole chain.

Use Cases

1. Developers building agents — Claude Code, Cursor, or your own: sessions are captured automatically, and the agent reads/writes memory via MCP.
2. Obsidian users — your notes stop being a static archive and become a knowledge base the agent maintains and queries.
3. Teams — the portable format means you can migrate memory between vendors without replaying history.