Field noteNov 20, 20265 min read1 source

Agent Memory. AI coding agents are great until they get stuck on the same bug, pile on broken code, and forget what they were building.

In one minute

AgentMemory is an open-source memory layer for coding agents that captures work on a developer’s machine and brings selected context into later sessions. Its project page describes integrations with 26 agents, keyless local search, and benchmark results; the practical question is whether cross-session recall can reduce repeated explanations without creating unacceptable data exposure.

What changed: AgentMemory introduces persistent, shared local memory for coding agents across 26 tools, including Claude Code, Cursor and Codex.

What happened

AgentMemory’s project page presents the software as persistent, searchable, shared memory for coding agents including Claude Code, Cursor, Codex, OpenClaw, OpenCode, and Hermes. It lists version 0.9.33, an Apache-2.0 license, 30.5K GitHub stars, and 681K all-time npm installs. The project says it works with 26 agents and has native plugins for 13, alongside MCP and REST access. Users can start it with npx -y @agentmemory/agentmemory@latest, then connect an agent with agentmemory connect <agent>. The stated requirement is Node.js 20 or newer.

The described workflow records supported agent activity through hooks or native integrations. Each event is stored as an observation, indexed for keyword and vector search, then related observations can be consolidated into memories, lessons, and graph facts. At the next session, recall selects context by project and agent and returns it within a user-controlled budget. The page says records link back to their observations, so a recalled memory can be traced to the session that produced it. On supported integrations, a PreCompact hook can restore memory context around compaction. The product also includes a local viewer for sessions and memories.

The page says capture, storage, keyword search, the REST API, and the viewer run locally by default. The bundled services bind to loopback addresses, and the runtime generates a secret on first start. No account or API key is required for keyless mode: BM25 and graph recall work without a model provider. Local embeddings can be enabled with EMBEDDING_PROVIDER=local; LLM-based consolidation requires configuring a provider and its key. The project says memory text stays local unless a remote service is configured, while configured remote LLM or embedding services receive the text they process. A configured remote Redis backend stores records remotely. It also says telemetry is disabled.

For version 0.9.33, the page highlights changes aimed at large stores and long operations. Listing records is now paginated; the release notes say this avoids a prior transport-size problem and supports scopes larger than 64 MiB. Saved-memory search is no longer capped at three results, and keyword and vector search each pass at least 40 candidates into fusion. The MCP shim waits 190 seconds for proxied calls, beyond the server’s stated 180-second limit. The page reports a keyless LongMemEval-S evaluation on 470 answerable questions: BM25 found every required source in the top five for 84.7%, versus 72.1% for a grep baseline; it found at least one relevant source for 96.2%. The page links a reproduction command and dataset details.

What the archive adds

We have not covered this before; it opens a new thread.

Trend signal

This supports New Runtime’s company memory needs write loops pattern: AgentMemory describes a cycle that captures work, consolidates observations into linked memories, and brings selected results back into later sessions. The page describes the mechanism, but does not establish that teams routinely correct or govern the resulting memory in production.

Why now

The project page identifies version 0.9.33 and calls out changes for large stores, broader recall, and long-running calls. Those changes matter as stored agent histories grow: a memory system that works only for small stores or short operations becomes less useful as teams accumulate sessions.

Why it matters

A shared memory layer could make switching between coding agents less costly: a developer may not need to re-explain project conventions or rediscover decisions made in another session. The implementation also makes the trade-off unusually visible. Capturing prompts, tool calls, and bounded output copies can create useful provenance, but it means the memory store may hold sensitive development context. Local-by-default operation, loopback binding, and optional model providers give teams control points; they do not remove the need to decide what may be captured, who can read the store, and whether any configured remote service is acceptable.

The benchmark is a narrower claim than “agents remember better.” It tests retrieval of required sources on a specified dataset, not whether a coding agent makes fewer mistakes, ships faster, or produces safer code. A team evaluating the tool should inspect recalled context and test it against its own workflows before relying on it to guide changes.

Limits and uncertainty

The benchmark figures are reported by AgentMemory, and the page does not establish independent replication or show that retrieval gains translate into better coding outcomes. Its headline adoption figures are project-page counts; they do not show how many users are active or how the software performs in production. Support also depends on integrations: automatic capture requires a native plugin or hooks, and the page notes that hook capture is opt-in for some connections. Local defaults do not guarantee local handling if a user configures remote storage, embeddings, or an LLM provider. The page does not quantify memory quality, false or stale recall rates, or the effort required to review sensitive captured data.

Skill to build now

Try a bounded evaluation before connecting every agent. Choose one non-sensitive project, connect one agent, and inspect what gets captured and recalled across two sessions. Check whether a recalled item points back to a useful source, whether stale instructions recur, and whether the returned context fits the task. Keep remote providers unset during this test if local handling is a requirement; the project says keyless keyword and graph recall work without an API key.

Reader takeaway

AgentMemory makes cross-agent, cross-session memory available as local-first open-source tooling, with keyless search and integrations for many coding agents. Its retrieval benchmark is promising but does not prove better coding outcomes. The decision for a team is therefore twofold: test whether recall helps in its own workflow, and treat captured prompts and tool output as data that needs a clear handling policy.

Evidence / sources

  1. 1agent-memory.dev
All Field Notes

Next reads

Read next

Open the graph
  1. 01topicCoding AgentsExplore the coding-agents topic hub.
  2. 02topicClaudeExplore the claude topic hub.
  3. 03topicClaude CodeExplore the claude-code topic hub.
  4. 04archiveField NotesOpen the latest editorial analysis.
  5. 05sourcesSourcesSee the original sources this archive cites.
Who read this page?Machine requests, hidden until opened

Loading the privacy-safe route aggregate...

Open the JSON contract