tigerless-labs/agent-memory

What it solves

AI agents typically lose all information once a session ends. agent-memory provides a long-term memory runtime that allows any agent (not just coding agents) to persist knowledge across sessions. It solves the problem of choosing between a complex retrieval engine (which is often opaque) and a simple filesystem (which doesn't scale or rank results), by combining both into a single store.

How it works

The system uses Markdown files as the single source of truth for memories, stored in a structured directory. A local SQLite index (FTS5) acts as a rebuildable cache for fast, ranked retrieval.

Key mechanisms include:

• Multi-track Retrieval: It uses a combination of deterministic injection (via MEMORY.md), BM25 recall over the index, and direct filesystem access (ls/grep).
• Retrieve by Path: Instead of pasting full text into the agent's context, it returns a list of file paths and abstracts. The agent then reads only the specific depth it needs (e.g., outline, abstract, or full file).
• Automatic Writing: Distillation of memories is triggered at conversation boundaries rather than relying on the agent's manual judgment.
• Sleep-time Management: A separate "sleep" process handles consolidation and proposes deletions, which the user must confirm.
• Host Agnostic: It does not include its own LLM client; it borrows the reasoning capabilities of the host agent's CLI (e.g., Claude Code or Codex CLI).

Who it’s for

Developers building or using AI agents that need persistent, portable, and human-readable long-term memory that can be shared across different agent hosts.

Highlights

• Files-as-Truth: All memories are Markdown files; the index can be deleted and rebuilt without any data loss.
• Zero API Keys: The library requires no API keys or external billing because it uses the host agent's existing CLI.
• Ranked Local Retrieval: Combines vector search (via plugin) and BM25 for precise recall without network calls.
• Cross-Host Compatibility: A single memory store can be shared between different agents like Claude Code and Codex CLI.