sapience
Mémoire sémantique persistante qui se cumule entre les sessions, plus un registre de jugement avec calibration réelle (scores de Brier, cartes de biais).
Documentation
Sapience
Human-like memory and a judgment ledger for AI — an MCP server for Claude Code.
An LLM has intelligence — it processes and analyzes brilliantly — but it's amnesiac between sessions and never accumulates your experience. Humans win on something else: memory that persists and judgment that gets sharper because we keep track of how our past calls turned out. That faculty — the one that makes Homo sapiens more than raw brainpower — is what Sapience adds to your AI.
Two halves:
- A human-like memory — episodic and semantic memories, ranked by importance, consolidated over time into durable patterns. Not RAG over a scratch file.
- A judgment ledger — log a prediction with a probability, resolve it against what actually happened, and get a real calibration read (Brier score, reliability by confidence band, a bias map) so you can see where your judgment is systematically off.
Sapience gives one user's AI a compounding memory + judgment loop. It is not a claim to reproduce human cognition — it's the missing feedback loop that lets an intelligence learn from experience.
The judgment ledger
This is the part you won't find in other memory tools. Every "AI memory" remembers what you said; Sapience keeps score of whether you were right.
- Log a forward-looking call with a probability (0–1) and — crucially — the reasoning and conditions as they were at the time. Most retrospectives rewrite history; this preserves the contemporaneous evidence.
- Resolve it when the outcome is known (right / partial / wrong).
- Calibrate. Sapience computes a Brier score against a base-rate baseline, breaks accuracy down by confidence band, and flags over/under-confidence. A Claude-written narrative sits on top of the numbers — never instead of them.
Honesty by design: below a sample threshold (20 binary-scored resolutions by default — partial resolutions don't count), Sapience refuses to call anything a "bias" and explicitly labels its output "reflection, not statistics." A bias is not a bias at n=3.
How the memory works
- Storage — a local ChromaDB vector store; the ledger is local SQLite. No third-party SaaS account.
- Embeddings — OpenAI (
text-embedding-3family) for semantic similarity. - Synthesis — Anthropic Claude for context briefs, consolidation, calibration, and bias maps.
- Retrieval — candidates are over-fetched by similarity, then reranked by
similarity × salience, so an important-but-slightly-less-similar memory can still surface.
Memory types: episodic (events/decisions), semantic (patterns, written by consolidation), user (facts about you), feedback (how to work with you), project (initiatives), reference (external pointers).
Privacy — read this precisely
Your data is stored locally (vector DB + SQLite on your machine; no hosted account). By default Sapience is not fully local compute: memory content is sent to OpenAI to create embeddings, and selected memories are sent to Anthropic for briefs, consolidation, and calibration. Embeddings can be made fully local with EMBEDDINGS_PROVIDER=local (a bundled MiniLM model — no key, no network after the first model download); briefs/consolidation/calibration narratives still require Anthropic. If that tradeoff doesn't work for your data, don't point Sapience at it.
Tools
Memory — search_memory, save_memory, get_context_brief, get_related, consolidate, list_memories, memory_stats
Memory admin — get_memory (inspect by id), edit_memory (fix content/salience/topic/type in place, re-embeds automatically), delete_memory, export_memories (JSONL backup), find_duplicate_memories (report-only — nothing is auto-deleted)
Judgment ledger — log_assessment (prefer a numeric probability), list_pending_assessments, resolve_assessment, generate_calibration (Brier + reliability, gated for sufficiency), get_bias_map
Setup
Requires Python 3.12+.
Install from PyPI:
pip install sapience-mcp
The PyPI distribution is named
sapience-mcp— PyPI's name-similarity rules blocked the baresapiencename — but everything else keeps the original name:import sapience, the installed command issapience, and the four console scripts (sapience,sapience-weekly-review,sapience-consolidate,sapience-demo) are unchanged.
Then create a .env in your project directory (variables below) or export them directly — Sapience picks up .env from your current working directory.
Or, from source (for development):
git clone https://github.com/allenc84/sapience.git
cd sapience
python3.12 -m venv venv
./venv/bin/pip install -e .
cp .env.example .env # then edit
Configure .env (see .env.example):
MEMORY_USER_CONTEXT="Jane Doe, founder of Acme" # who the memory serves
OPENAI_API_KEY=sk-proj-...
ANTHROPIC_API_KEY=sk-ant-...
# Optional:
LEDGER_DOMAINS="predictions,decisions,commitments" # your judgment domains
SAPIENCE_DATA_DIR=/absolute/path/to/data # defaults to a per-user OS dir
SAPIENCE_NAMESPACE=work # memory namespace (default: "default")
EMBEDDINGS_PROVIDER=openai # or "local" (bundled MiniLM, no key needed)
EMBEDDINGS_MODEL=text-embedding-3-small # OpenAI model when provider is openai
Switching embedding providers on an existing database requires re-embedding everything (dimensions differ). With the server stopped:
EMBEDDINGS_PROVIDER=local python -m sapience.repair --rebuild --re-embed --server-stopped
Namespaces
Memories are partitioned by namespace — set SAPIENCE_NAMESPACE per project/workspace (e.g. in a project's .mcp.json env block) to keep contexts separate inside one database. Reads and writes default to the server's namespace; pass namespace: "*" to search_memory/list_memories to read across all of them, and memory_stats shows the per-namespace breakdown. Records created before namespaces existed are stamped default automatically on first read. The judgment ledger is deliberately not namespaced — your track record is yours, not a project's.
macOS Keychain (optional): the
run_*.shscripts read keys from the Keychain if present, falling back to.env. Store keys as the-wargument, never via the interactive prompt — the prompt truncates at 128 chars and silently corrupts longer keys:security add-generic-password -U -s "OPENAI_API_KEY" -a "claude-memory" -w 'sk-proj-...'
Install as a Claude Code plugin (easiest)
With uv installed and OPENAI_API_KEY + ANTHROPIC_API_KEY in your environment:
/plugin marketplace add allenc84/sapience
/plugin install sapience@sapience
This wires up everything below in one step: the MCP server (launched via uvx, no manual install) and the /sapience:log judgment-ledger command. Consolidation and the weekly ledger review run inside the shared server, so there are no hooks or cron jobs to set up. Configuration still comes from your environment — set MEMORY_USER_CONTEXT, LEDGER_DOMAINS, SAPIENCE_DATA_DIR, or SAPIENCE_PORT there if you want non-defaults.
Wire into Claude Code manually
Add to your MCP config (~/.claude.json or project .mcp.json):
{
"mcpServers": {
"sapience": {
"command": "/absolute/path/to/sapience/run_server.sh"
}
}
}
Or, with the package installed, point directly at the console script / module:
{ "mcpServers": { "sapience": {
"command": "/absolute/path/to/sapience/venv/bin/python",
"args": ["-m", "sapience.server"],
"env": { "SAPIENCE_DATA_DIR": "/absolute/path/to/data" }
} } }
Restart Claude Code. The server reads keys and config at launch — restart after changing either.
How it runs: one shared server, any number of sessions
Chroma's on-disk vector index is single-writer: two processes holding the same chroma_db each flush their own copy of the index and overwrite each other, and the index silently drifts until a server segfaults on open. So Sapience never runs one server per session.
What each Claude Code session launches (sapience, run_server.sh, or python -m sapience.server) is a thin stdio relay. It looks for the shared server on 127.0.0.1:8765 (SAPIENCE_PORT to change), starts it in the background if none is up, and forwards the session's MCP traffic to it. The shared server:
- is the only process that opens the DB, enforced by an exclusive lock (
chroma_db.lock) — anything else that tries fails fast withAnother sapience process already owns ...instead of corrupting the index; - binds to localhost only and is stateless, so sessions keep working across a server restart — if it dies, the next request from any session restarts it;
- runs consolidation for every namespace every 6 hours (
SAPIENCE_CONSOLIDATE_EVERY_S) and checks whether the weekly ledger review is due every hour; - logs to
server.login the data directory, and is replaced automatically when a relay from a different installed version connects. Per-projectSAPIENCE_NAMESPACEstill works: each relay applies its own namespace to its calls.
Nothing to configure. If you'd rather keep the server under your own supervisor (launchd, systemd), run sapience --http there; relays will use it as-is. sapience --direct is the old in-process stdio server — single session only, since it holds the lock. Stop the server before python -m sapience.repair --rebuild; the lock refuses otherwise.
The /log command
.claude/commands/log.md provides a /log slash command for the ledger — logging, reviewing, resolving, and generating calibrations/bias maps in natural language. Copy it into your project's .claude/commands/.
Automation (optional)
Both jobs run inside the shared server and need no setup. The scripts remain for one-off manual runs while the server is stopped (each opens the DB and will hit the lock otherwise):
run_consolidate.sh— nightly: extract semantic patterns from recent episodes (cron/launchd).run_weekly_review.sh— weekly ledger review; designed for a Claude Code Stop hook.
Try it on demo data
Don't want to point Sapience at real data yet? Seed a fictional founder's dataset — 21 memories and a 30-call judgment ledger with a real calibration story for the bias map to find (overconfident on product bets, calibrated on hiring, underconfident on growth):
OPENAI_API_KEY=... sapience-demo --dir ./sapience-demo-data
It prints the MCP config to paste, plus a 4-step demo flow. Everything is fictional; the target directory must be new or empty.
Migrating existing markdown memories
MEMORY_MIGRATE_DIR="$HOME/path/to/memory" ./venv/bin/python -m sapience.migrate
License
MIT — see LICENSE.