agent-memory gives coding agents local long-term memory

What this is about

agent-memory is a new open-source tool from Tigerless Labs that gives Claude Code, Codex CLI, and other command-line agents a shared long-term memory. The GitHub project was created on September 1, 2026 and uses the MIT license. Instead of keeping memories only in a proprietary database or remote service, it treats ordinary Markdown files as the authoritative source.

This is mainly useful for people who let several agents work on the same projects over extended periods. Decisions, corrections, and project knowledge can survive across sessions without requiring a separate storage service or API key. On October 5, 2026, the project was still young: its repository identified version 0.1.0, and the installation guide said that no PyPI release was available yet.

What agent-memory actually does

The tool stores each memory as a Markdown file in a local directory. A central MEMORY.md acts as a compact index. Alongside it, agent-memory builds a disposable SQLite index for full-text search. A query can first return short results with a file path and score; the agent then opens only the material it actually needs. This avoids copying the entire knowledge store into the model context for every question.

Command-line operations can record, retrieve, correct, supersede, merge, and time-scope memories. An optional MCP interface exposes similar functions to compatible agents. The repository provides setup commands for Claude Code, Codex CLI, and Muse Code. A background management pass is intended to consolidate older entries and propose deletions rather than execute them without approval.

The architecture deliberately separates original files from the search index. If the index is removed, it should be rebuildable from the Markdown files. That separation improves portability and auditability: users can read, search, version, and back up the files with ordinary tools.

Why it matters

Agents often lose working context between sessions. Teams compensate with long startup instructions, manual handovers, or ever-larger context windows. agent-memory takes a different route: knowledge persists outside the model and is retrieved selectively when needed. This could save time for architectural decisions, recurring mistakes, and long-running software projects.

The approach also matters for privacy and switching providers. As long as the store remains local, the material is not automatically entrusted to another SaaS vendor. Because Markdown is authoritative, users can inspect and migrate it without specialized software. That does not mean content never leaves the device: when a cloud-based agent reads a retrieved file, its contents may still be sent to that model provider.

The repository reports its own measurements on a bounded variant of LongMemEval-S. Those numbers are not an independent product evaluation. LongMemEval itself was introduced in 2024 as a benchmark for long-term memory in chat assistants and provides useful context, but the experiment conditions described by this project differ from published standard results.

In plain language

Think of agent-memory as a carefully maintained workshop binder. Experience is kept on readable sheets, while a card index points to the sheet containing a particular decision. If the card index breaks, the sheets remain intact and a new index can be built from them.

A practical example

A team operates three services and uses both Codex CLI and Claude Code. Across 20 sessions it accumulates 60 relevant decisions: database migrations, rejected libraries, and known production traps. Without shared memory, the team repeatedly explains the same rules.

With agent-memory, it stores every decision as a Markdown file. When an agent later asks about the authentication strategy, local retrieval first returns eight short candidates. The agent opens two relevant files in full and learns why an earlier solution was rejected. A human can inspect the same files in an editor, version them in Git, and trace a correction if a memory was wrong. The sensible first test is a disposable project with a handful of synthetic decisions, not the production knowledge archive.

Scope and limits

First, agent-memory was an early project on October 5, 2026. Installation from a Git checkout, Python 3.12 or newer, and uv require technical experience; a convenient package and broad product integrations are not yet available.

Second, local storage does not solve the quality problem by itself. An agent can preserve false, duplicate, or sensitive information. Teams therefore need rules for write access, corrections, retention, and secrets. Credentials in particular do not belong in ordinary Markdown memories.

Third, the published performance claims come from the project itself. The README referenced an experiment document that was not publicly reachable at the stated path on October 5, 2026. Until independent reproductions exist, teams should measure the claimed benefits in their own workflow.

SEO & GEO keywords

agent-memory, Tigerless Labs, coding agents, long-term memory, Claude Code, Codex CLI, local AI, Markdown, SQLite FTS5, MCP, open-source AI, agent memory