Local-first memory for AI agents

Your machine. Every agent. One brain.

Sigil gives Claude Code, Codex, Cursor, Kiro and any MCP client one shared memory. It captures what you decide as you work and recalls it in milliseconds, from your own Postgres. No cloud. No telemetry.

Get started in 5 minutes → Watch it work ↓
$curl -fsSL https://raw.githubusercontent.com/Anmol-Srv/sigil/master/install.sh | sh
Retrieval R@10
100%
p50 search
33ms
Tokens per prompt
~1.5K
Config files to edit
0

01 / 05Amnesia

Every session starts at zero.

Each agent keeps its own walled memory, so you re-explain the same context every day.

02 / 05Capture

You never write a remember call.

Four Claude Code hooks capture decisions as you work. Other agents write through the CLI or MCP.

03 / 05Recall

Ask in plain words. Get the right fact.

pgvector + BM25 hybrid search, re-ranked by recency and co-recall. 33ms p50.

04 / 05Agents

Every agent reads the same brain.

sigil init wires Claude Code, Codex, Cursor and Kiro. Any MCP client takes one line.

05 / 05Sigil

Nothing leaves your machine.

Your own Postgres. No cloud, no telemetry. Fully air-gappable with Ollama.

Capabilities

Infrastructure, not a feature.

A background daemon, your Postgres, and a memory model built for agents. Hover a number to trace it through the system.

Claude Code 4 hooks · auto Codex CLI MCP + rules Cursor MCP + rules Kiro MCP + steering Any agent shell CLI · MCP sigil daemon localhost capture hooks · cli · mcp · dedup hybrid search pgvector + bm25 → rrf re-rank act-r + hebbian pods project · person · playbook your Postgres on your machine facts + embeddingsfull-text indexknowledge graph home server sigil · same Postgres LLM ollama or cloud embedder ollama or cloud 01 02 03 04 05 06 07 08
  1. 01 Universal CLIAnything with a shell can search and remember.
  2. 02 Auto-captureFour hooks save decisions as you work.
  3. 03 Shared memoryEvery agent reads and writes one brain.
  4. 04 Cross-machineTwo machines, one Postgres, one memory.
  5. 05 Hybrid searchVector + keyword, fused in one query.
  6. 06 Re-rankingRecency, frequency, co-recall.
  7. 07 Pluggable podsNew pod kinds are contract files.
  8. 08 Air-gappableOllama for both models. Nothing leaves.

Works with

Wired into every agent you already use.

sigil init detects and configures the first four automatically. The rest take one line from sigil register --print.

zsh — ~/api

$sigil init

scanning for coding agents…

✓Claude Code4 hooks + mcp~/.claude/settings.json

✓Codex CLImcp + rules~/.codex/config.toml

✓Cursormcp + rules~/.cursor/mcp.json

✓Kiromcp + steering~/.kiro/settings/mcp.json

◆memory wired into 4 agents. open one and start working.

$sigil register --print

Add this to your MCP client config:

{
  "mcpServers": {
    "sigil": { "command": "~/.sigil/bin/sigil-mcp", "args": [] }
  }
}

$

Auto-wired by sigil init

Claude Code4 hooks + mcpCodex CLImcp + rulesCursormcp + rulesKiromcp + steering

One line from sigil register --print

ContinuemcpClinemcpWindsurfmcpClaude DesktopmcpCustom agents & CImcp

Honest benchmarks

Public, reproducible. No corpus theater.

Measured on LongMemEval, with the caveats printed next to the numbers.

Recall LongMemEval oracle · n=100
100% R@1
100% R@3
100% R@10
Search latency avg / p95
33msavg
61msp95

One SQL query: pgvector + BM25 fused by Reciprocal Rank Fusion.

Cost per full run

$0.21

100 questions, OpenAI top-quality stack. Free on Ollama.

Caveats: oracle is the easy split, n=100 is small, and each haystack is ~25 chunks. Scripts and methodology: eval/longmemeval.

Install

Get a working brain in one sitting.

  1. 01InstallPuts Sigil on your PATH and opens the dashboard.
  2. 02Pick providersAn LLM and an embedder, with live connection tests.
  3. 03Open your agentMemory is already wired in.
$curl -fsSL https://raw.githubusercontent.com/Anmol-Srv/sigil/master/install.sh | sh
Read the quickstart → Star on GitHub