Memory

The job

Many agents start each session without what earlier ones learned. A memory tool keeps something from a session, such as decisions, facts, notes or a map of the code, and offers it back later.

How they differ

One kind records coding sessions through hooks and replays the useful parts later. Agentmemory, claude-mem, ai-memory and Trail-Code work this way, and atlas adds a team timeline that links sessions to commits. Ai-memory keeps notes as git-backed Markdown files. These may suit people who move between assistants and want a single record of past work.

A second kind builds a graph or index of a codebase for agents to query. Codebase-memory-mcp and graphify do this for code, and semantica builds knowledge graphs with recorded provenance for developers to put beneath an agent framework.

A third kind is a memory layer to build into your own application. Mem0 is offered as a library, a server and a hosted platform. Hindsight keeps memory in banks per user, agent or project, separated by metadata the integrator sets. Gbrain keeps facts with their sources and lets agents revise them. Xerj is a search engine with a memory store, and utopia keeps dated facts in a graph where, by its account, agent writes arrive as proposals. OpenLore serves a folder of Markdown with access grants. These suit developers who need long-term memory for chatbots or agents they are building.

A fourth kind learns skills or saves cost. Autoharness turns finished sessions into reusable skills for Claude Code. TokenEconomizer learns from earlier sessions to cut token spend. Hermes-agent keeps memory and writes skills as part of a general agent.

A fifth kind is a whole agent platform in which memory is one part. KiroCrew, rakazo and munder-difflin are examples. Ruflo and loopx instead add coordination and lasting state to existing agents. Webcmd keeps a local memory of websites an agent has visited, and OpenResearch tracks experiments in a local database. They may suit people who want the agent and its memory from one place.

Licences

On 10 November 2026, GitHub detected Apache-2.0 or MIT on most of the tools here. NanoMuse carried GPL-3.0. A licence can change, and each tool's own page gives the latest reading with its date.

What to check first

Many of these tools capture what an agent does. The agentmemory README says hooks record session activity automatically, and its default scope lets each connected agent see what the others saved. Ai-memory records observations through lifecycle hooks. Atlas records every session when capture is on, and its secret redaction is the project's own claim. Trail-Code says setup reads earlier Claude Code sessions for the repository and commits team learnings. KiroCrew tasks can run unattended, and munder-difflin's coordinating agent settles routine requests itself.

Some send content off the machine. Graphify says documents, images and video are handled by a model outside the tool. Gbrain says text may reach cloud providers you configure, and warns that an unrelated npm package with its name can replace the binary. NanoMuse sends conversation text through the developer's relay unless you host one, and can drive phone apps and windows by screen. Claude-mem, OpenResearch, KiroCrew and munder-difflin describe usage analytics or telemetry.

Webcmd examples use logged-in browser profiles. Hermes-agent can import allowlisted API keys from other agent tools when asked. Octop writes admin credentials to a text file on first Docker boot, by its README's account. TokenEconomizer, autoharness and ruflo edit agent settings or write files into the project or home directory, and ruflo says its plugin mods run with the account's permissions and no sandbox.

Early or experimental status is stated by utopia, which names an rc7 release, by xerj, which names rc.101, and by semantica, which lists breaking changes. Mem0 has a flag that turns server authentication off, xerj's examples start its server with an --insecure flag, and OpenLore's quick start opens SSH and web ports.