Back to projects
Active

Anamnesis

Persistent, distributed memory substrate for AI agents, engineered so no single machine is required to write or refresh it.

PythonSQLiteSigstore / CosignMCPGemini EmbeddingsGitHub Actions
GitHub RepositoryRead Architecture & Manifesto

Persistent memory for AI agents, designed from the ground up so that losing any single laptop or server never disrupts the memory pipeline. The memory pillar of Stewardship Engineering.

The Problem

Most AI memory implementations quietly depend on a single machine or hosted database instance. If that machine sleeps, reboots, or crashes, the agent loses context or fails to record new learnings.

Furthermore, traditional vector stores decouple memory storage from version control. When an agent hallucinates or stores incorrect facts, tracking who introduced the error and rolling it back is nearly impossible.

Architecture: Zero Single-Point-of-Failure

Anamnesis splits memory operations into three asynchronous phases:

WRITE (any device, offline-ok)       BUILD & SIGN (serverless CI)         SERVE (disposable replicas)
  memory = git commit to data/   ─►   GitHub Actions runner         ─►    pull latest Release
  (laptop, server, agent, phone)      index + embed + Cosign sign         verify signature ─► atomic swap
  1. Write Path: Memory entries, decisions, and feedback are plain Markdown files committed to Git. Writes work offline, merge cleanly, and maintain a permanent cryptographic commit history.
  2. Build & Index Path: GitHub Actions acts as the serverless build farm. When changes land on main, CI embeds entries, generates a unified SQLite database (memory.db), and cryptographically signs the release artifact using keyless Sigstore/Cosign.
  3. Serve Path: Any machine running an agent pulls the signed artifact, verifies the Cosign signature against the repository workflow identity, and atomically swaps the database into service.

Security: Preventing Belief Breaches

When an agent's memory database is served as a downloadable artifact, a compromised write token could attempt to publish a poisoned memory database. The agent would read the poisoned memory, believe it, and act on corrupted instructions.

To prevent belief breaches, every release is signed via Cosign. Local replicas verify that the artifact was generated strictly by the trusted repository workflow on main before accepting the update.

Serving to Agents via MCP

Anamnesis integrates directly with developer workflows via the Model Context Protocol (MCP). Agents in Claude Code, Codex CLI, or custom harnesses call recall, session_context, and get_decisions as native tool calls, retrieving relevant context in sub-50ms queries.

Open Source & Reading