Skip to content

Memory for coding agents

A memory layer for AI coding agents carries what an agent learned or did into its next session. Coding agents already include one kind: instruction files such as CLAUDE.md and AGENTS.md, plus notes Claude Code and Codex write for themselves. Memory servers and SDKs add facts that persist across tools or inside apps you build.

Hansard is a local-first memory layer that keeps the full record instead. It archives every conversation and memory file from your coding agents on your computer, and every agent you connect can search it.

Which memory layer is best for AI coding agents?

Section titled “Which memory layer is best for AI coding agents?”

No single memory layer is best for every setup. Choose by what you want remembered and where you allow it to be stored. Most setups combine two or three of these:

  • Rules for a repository: start with instruction files. AGENTS.md at the repository root is read by Codex, Cursor, GitHub Copilot, Amp, and Devin, and by Claude Code when the folder has no CLAUDE.md. Instruction files cost nothing and travel with the repository.
  • Preferences one agent learns: turn on that agent’s own memory, such as Claude Code auto memory or Codex Memories. Both stay on your computer and serve only that agent.
  • Facts several agents read and write: use a memory server such as Mem0, Supermemory, claude-mem, agentmemory, Basic Memory, or Cognee. Check where each one stores your data, since some run locally and some need a hosted account.
  • Memory inside an application you build: use an SDK such as Mem0, Letta, Zep, Graphiti, or Cognee.
  • Everything your agents did, in every harness: use Hansard. It keeps whole conversations rather than extracted facts, so an agent can read what was tried, decided, and reverted in the original messages, whichever harness did the work.

These kinds work together. Hansard backs up the instruction files and built-in memories your agents keep, and a memory server can run beside it.

Kind What it keeps Where it lives Examples
Instruction files Rules you write The repository and your home folder AGENTS.md, CLAUDE.md, Cursor project rules
Built-in agent memory Notes an agent writes for itself Your computer Claude Code auto memory, Codex Memories
Memory servers Facts or notes extracted from sessions Your computer or a hosted service, depending on the tool Mem0, Supermemory, claude-mem, agentmemory, Basic Memory, Cognee
Memory SDKs Memory inside an app you build Your infrastructure or the vendor’s cloud Mem0, Letta, Zep, Graphiti
Conversation archive with recall Every conversation and memory file, as written Your computer Hansard
Option What it stores Where the data lives License Coding agents
Claude Code memory CLAUDE.md instructions, and auto memory notes. The first 200 lines or 25 KB of MEMORY.md load at each session start. Local files. Auto memory is not shared across computers. Part of Claude Code Claude Code
Codex memory AGENTS.md instructions, up to 32 KiB combined by default, and optional Memories summarized from earlier chats Local files under ~/.codex Part of Codex Codex
Cursor rules Project, user, and team rules, plus AGENTS.md. Cursor’s documentation no longer describes Memories. The repository, Cursor settings, and the team dashboard Proprietary Cursor
Mem0 Facts a model extracts from conversations The library stores memories locally and sends text to a model provider, OpenAI by default. Its coding-agent plugins and hosted MCP server keep memories in a Mem0 account. Apache-2.0, with a hosted platform Claude Code, Codex, and Cursor plugins, and a hosted MCP server
Supermemory Extracted facts, profiles, documents, and connector data A hosted service by default, or a local server whose source is not public MIT repository, with hosted plans Plugins for Claude Code, Codex, Cursor, OpenCode, and others
claude-mem Observations and summaries a second model writes from each tool call ~/.claude-mem, with opt-in cloud sync Apache-2.0, with paid hosted features Claude Code, Codex, Cursor, OpenCode, and others
agentmemory Observations, session summaries, facts, and workflows A local server Apache-2.0. It runs on iii-engine, licensed Elastic-2.0. Claude Code, Codex, Cursor, and others through plugins, hooks, and MCP
Basic Memory Markdown notes linked into a knowledge graph ~/basic-memory, with an optional hosted cloud AGPL-3.0 MCP for Claude Code, Codex, Cursor, and others
Cognee A knowledge graph of documents, code, and conversations Local by default, self-hosted, or Cognee Cloud Apache-2.0, with a hosted cloud Claude Code and Codex plugins, and MCP for Cursor and others
Letta Agent memory as a git-versioned folder of Markdown files Letta Cloud by default, or local Apache-2.0, with hosted plans Letta’s own coding agent, not a plugin for other agents
Zep and Graphiti A knowledge graph of entities and facts that tracks when each was true Zep Cloud. Graphiti is self-hosted. Zep is proprietary. Graphiti is Apache-2.0. Zep’s MCP server for Claude Code, Codex, and Cursor. Graphiti has a reference MCP server.
Hansard Every conversation as written, with tool calls, plus revisions of memory and instruction files ~/.hansard on your computer MIT Recall over MCP in Claude Code, Codex, Cursor, and more than a dozen other agents. Imports history from dozens of harnesses.

Each row was checked on 2026-11-08 against the project’s own documentation, README, or license file. Plugin lists and plans change often, so follow the links under Sources for current details.

How do I give Claude Code and Codex the same memory?

Section titled “How do I give Claude Code and Codex the same memory?”

Share the rules through AGENTS.md, and the history through Hansard recall.

  1. Put shared rules in AGENTS.md at the repository root. Codex reads it. Claude Code reads it when the folder has no CLAUDE.md. If you keep a CLAUDE.md, add the line @AGENTS.md to it, or type /config in Claude Code and set Project instructions to claude-md-and-agents-md.

  2. Install Hansard recall in both agents:

    Terminal window
    hansard integrations add-to claude
    hansard integrations add-to codex
  3. Ask either agent about earlier work. It searches both agents’ conversations with search_past_sessions, and both agents’ memories with search_memories, so a Codex session can use a preference Claude Code recorded.

  4. To merge duplicate memories and settle contradictions between the agents, run /unify-memories in Claude Code, or ask Codex to unify your memories.

See Recall in your agents and Memories.

  • Imports conversations from dozens of harnesses, and keeps a copy of each original file, so history stays after a harness deletes its own. See Supported harnesses.
  • Backs up memory and instruction files every 15 minutes while the watcher runs, with a revision history of each file.
  • Gives agents read-only recall tools over MCP: search_past_sessions, get_past_session, search_memories, get_memory, code_history, and work_context. See Recall tools.
  • Drafts lessons from what failed and what was fixed once you turn on dreaming, and adds a lesson to a file only after you approve it. See Review proposed lessons.
  • Runs on your computer without an account. Search by meaning uses a local embedding model that you turn on.
  • Hansard is not an SDK or a hosted memory API for apps you build. Use a memory SDK for that.
  • Hansard does not add remembered facts to every prompt. An agent searches the archive when it needs earlier work.
  • Hansard keeps conversations and memory files, not a separate store of extracted facts.
  • A Team shares conversations, not memories. Memories stay on each computer. See Teams.
  • Chats in a browser, such as ChatGPT on the web, cannot call the tools installed on your computer. They reach your archive through the AI assistants connection, which works while Hansard is running on your Mac.
  • What an agent reads from the archive goes to that agent’s model provider, like any other context. See What leaves your machine.