
agent-memory-bridge
io.github.zzhang82/agent-memory-bridge
Local MCP memory for coding agents: store decisions with their reasons, recall them on request.
Documentation
Agent Memory Bridge
Teach one decision and its reason. The next session can recall that reason when it recalls at the start.
AMB keeps a confirmed project decision in a local store. A later session gets it back by recalling, not by rereading the old chat.
The loop
One process stores a decision. A second process, on the same AMB home, recalls the claim and the reason. The recording is that local stdio proof. It is not a filmed coding-agent session, and it is not an adoption measurement.
The source is examples/demo/first-win-demo.cast. The block below is an abridged transcript of that cast. Regenerating it runs the two processes, then writes generated playback: the header time and the line delays are fixed, not a measurement of the run. Command: python ./scripts/build_first_win_cast.py. Rendering a GIF needs agg; this repository does not include that GIF.
# Session A: store one explicit decision
namespace: project:checkout-app
claim: Merge pull requests only after CI is green on the target branch.
reason: Broken main blocked two releases this month.
stored: true
# Session B: a second stdio process, same home, recall at start
query: What is required before we merge a pull request?
claim: Merge pull requests only after CI is green on the target branch.
reason: Broken main blocked two releases this month.
same_record: true
fresh_process: true
Recall at the start
Connecting the server is not enough. The next session has to recall on purpose. This page does not claim that connecting MCP, or leaving a skill where it can be found, makes that happen by itself. There is no automatic learning.
Put this rule where the next session will actually follow it:
Before changing code, recall this project's Agent Memory Bridge decisions and their reasons.
Use the namespace from `project resolve`. Call the public `recall` tool.
Store a decision only after it is confirmed. Do not treat chat history as already stored.
Install, then bind the project
AMB requires Python 3.11+ and Git.
Current package/source version: 0.39.1.
python3 -m venv .amb-venv
.amb-venv/bin/python -m pip install agent-memory-bridge==0.39.1
.amb-venv/bin/python -m agent_mem_bridge project init . --yes
.amb-venv/bin/python -m agent_mem_bridge doctor
In later commands, <venv-python> means .amb-venv/bin/python on Linux and macOS. project init detects the Git repository and proposes a namespace such as project:my-app; it accepts --yes for non-interactive runs and --dry-run to preview without writing. It does not learn decisions from the tree. After that, project resolve . returns the namespace.
Run <venv-python> -m agent_mem_bridge with one persistent AGENT_MEMORY_BRIDGE_HOME. Configure a client with connect <client> --yes (and triggers install <client> --yes), or write the command shape by hand from Integrations. Connection is real when the client lists store and recall.
Store one confirmed decision with store:
record_type: decision
claim: Merge pull requests only after CI is green on the target branch.
reason: Broken main blocked two releases this month.
Free-form kind="memory" notes still recall, while explore and inspect list records that declare record_type: decision or record_type: constraint. In the next session, use the startup rule above and ask for that decision.
What you can do with AMB
- Decisions and Rules: Teach one confirmed project decision and its reason so the next session recalls why the code is written that way instead of guessing from chat history.
- Replacing Outdated Decisions: When a decision changes, replace the older record cleanly so outdated rules stop surfacing by default while remaining in change history.
- Agent Handoffs and Claim Timeouts: Post a short-lived handoff item that one worker claims with a timeout so parallel agents do not step on the same task.
- Tracked Work Runs and Proof Receipts: Record multi-step run progress, verified outcomes, and tamper-evident receipts showing which memories surfaced.
- Connected Clients and Trigger Reminders: Configure Claude Code, Cursor, Codex, Cline, and VS Code with a three-tier connection check (
config_valid,server_spawn_verified, andhost_cli_verified/host_reload_required), and install per-client reminders for the six workflow trigger moments.
Reading the Python source or raw MCP JSON? See the Plain-Words to Code Glossary for the exact code and wire names (namespace, supersedes, signal, lease_seconds, episode, attestation).
Clients, triggers, an existing home, and optional review
Connect a supported client in one command and run the three-tier connection check:
<venv-python> -m agent_mem_bridge connect claude-code --yes
<venv-python> -m agent_mem_bridge connect cursor --yes
<venv-python> -m agent_mem_bridge connect codex --yes
<venv-python> -m agent_mem_bridge connect cline --config-path ~/.cline/data/settings/cline_mcp_settings.json --yes
<venv-python> -m agent_mem_bridge connect vscode --yes
<venv-python> -m agent_mem_bridge connect cursor --check
connect writes .mcp.json (Claude Code project scope, or VS Code Copilot project format), .cursor/mcp.json, .codex/config.toml, .vscode/mcp.json, or Cline settings (cline_mcp_settings.json). Point a client at a remote HTTP authority with connect <client> --url <url> --token-file <path> --scope user. Project-scope token writes require --allow-project-token, and that config file must stay gitignored and must not be committed.
Install per-client trigger reminders and inspect the honest 6×5 client support matrix (native_hook, deferred_hook, trust_gated_hook, instruction_rule):
<venv-python> -m agent_mem_bridge triggers matrix
<venv-python> -m agent_mem_bridge triggers install claude-code --yes
<venv-python> -m agent_mem_bridge triggers install cursor --yes
<venv-python> -m agent_mem_bridge triggers test claude-code
triggers install writes .claude/settings.json, .cursor/hooks.json, .codex/hooks.json, AGENTS.md, .github/hooks/amb-triggers.json, .github/copilot-instructions.md, .clinerules/amb-triggers.md, and .cline/rules/amb-triggers.md. New untracked project files from connect or triggers install mark the Git worktree dirty. explore and inspect report dirty_worktree until those files are committed or gitignored.
If an existing bridge lives at ~/.codex/mem-bridge, v0.36+ does not open it automatically. Set AGENT_MEMORY_BRIDGE_HOME to that directory, or copy it to ~/.local/share/agent-memory-bridge. The configuration guide has the cutover.
The companion skills are optional help for discovery and everyday use. They do not replace the startup rule above; see trigger moments and hook recipes for additional background. Install them from Agent-Memory-Skills:
npx skills add zzhang82/Agent-Memory-Skills --skill amb
npx skills add zzhang82/Agent-Memory-Skills --skill amb-connect
amb(raw)amb-connect(raw)
For a source checkout, install with <venv-python> -m pip install -e . instead of the pinned package.
Human-first Explore answers “What does AMB currently know?” Inspect answers “Why did this information surface?” Both are local and read-only. They are not the success check.
<venv-python> -m agent_mem_bridge explore --namespace project:my-app
<venv-python> -m agent_mem_bridge inspect \
--namespace project:my-app \
--query "What is required before we merge a pull request?"
Code tells AMB what the repository contains. A confirmed decision tells it why. Derived repository facts come from a clean Git commit. If HEAD moves or the worktree is dirty, refresh on purpose:
<venv-python> -m agent_mem_bridge bootstrap-repo . --namespace project:my-app
Refreshing repository facts leaves stored decisions unchanged. Explore is CLI-only. It is not an extra MCP tool.
doctor and verify check server prerequisites. They do not prove that a client loaded its MCP config. first-run is optional guided help, not the first-win path. Historical Codex observations from the v0.35 line are in First-win acceptance.
How a decision stays traceable
The useful part is not remembering more. It is knowing where a decision came from, whether it is still current, and how it changed.
| Memory concern | AMB approach |
|---|---|
| Current repository truth | Derived from a clean repository state and refreshed explicitly |
| Human decisions and constraints | Stored explicitly as governed durable memory |
| Changed knowledge | Revised or superseded instead of silently overwritten |
| Why context surfaced | Inspectable through local derived views and evidence paths |
| Cross-session reuse | Shared through the same configured local AMB home |
| Privacy | Local-first; no hosted memory service is required |
Code tells the next session what exists. AMB retains the confirmed reasons the code does not explain.
What AMB is — and is not
AMB is a governed local project-memory layer for coding agents. Durable knowledge, derived repository facts, provenance, and corrections stay distinguishable.
It is not a transcript archive, a promise that an agent will remember everything, or a system that silently converts every conversation into durable truth. There is no automatic learning.
Memory hygiene: Structure over prose
Do not treat memory as a prompt trash can for linter errors, directory conventions, or syntax dos and don'ts. If an agent makes the same mechanical mistake twice, harden the repository environment with linter rules, type guards, or pre-commit checks. Reserve AMB durable memory for the reasons that code cannot convey: architectural decisions, domain invariants, and third-party library gotchas. See the Memory Hygiene Guide for the 3-question ingestion filter.
Want the details?
| Read | For |
|---|---|
| Glossary | Plain-words feature names mapped to exact code, JSON, and MCP wire terms |
| First-win demo | The local two-process recording, and the older coordination demo |
| First-win acceptance | Historical Codex observation from the v0.35 line |
| Architecture | System shape and data flow |
| Authority model | Durable authority, derived views, correction, and audit rules |
| Knowledge Explorer | Human-first read-only project view |
| Memory Hygiene | Anti-pollution rules and the structure-over-prose boundary |
| Production Status | Current implementation facts, evidence, and known limits |
| Integrations | Generic stdio MCP installer contract |
| Install for Agents | Full install-to-first-success workflow |
| Configuration | Complete configuration reference |
| Examples | Sanitized demos and artifacts |
Technical model
Internally, AMB keeps code facts read from Git (derived_repository) separate from human-confirmed decisions (governed_durable_memory) so one cannot silently become the other. See the Plain-Words to Code Glossary for the exact code and wire terms. For maintainers and reviewers, the current authority flow is:
flowchart LR
A[Decisions & Rules / WHY] --> C[Current-Version Recall]
B[Code Facts / WHAT] --> D[Context Compiler]
S[Runtime State] --> D
C --> E[Task Memory]
E --> D
D --> F[In-Memory Context]
F --> G[Context Proof Receipt]
G --> H[Tracked Work Runs]
H --> I[Verification Receipt]
I --> J[Verified Outcome]
SQLite/WAL rows are durable authority. Repository snapshots, FTS rows, embedding sidecars, compiled context, reports, and Explorer views are derived. Context bodies are rendered in process and are not durably persisted by the compiler.
Optional HTTP gateway & deployment
AMB runs as a local-first stdio process by default, with an optional official Streamable HTTP transport and Optional HTTP gateway:
- Local stdio: no authentication or gateway configuration; it runs as before for individual agents.
- Optional HTTP gateway: per-client Bearer tokens, 4 roles (
admin,writer,reader,orchestrator), namespace limits (project:*,domain:*), and structured JSONL audit lines. - Audit kept out of SQLite: audit lines go to stdout, stderr or a file, not the database.
For remote server, NAS, or container setup, see Remote Deployment.
Trust and privacy
AMB is local-first. It does not require a hosted memory service. It separates durable memory from coordination Signals and mutable Dynamic State, keeps provenance visible, and rejects raw transcripts, hidden reasoning, and inline artifact bodies from the durable episode path.
Read the Trust Boundary, Authority Contract, and Closed-Loop Episode Authority for the exact boundaries.
MCP Tools
AMB exposes 17 public MCP tools:
store,recall,browse, andstatsforget,feedback,promote,annotate,revise, andexportbegin_run,record_run_event,get_run, andcomplete_runclaim_signal,extend_signal_lease, andack_signal
The public tool surface stays small. Setup, Project Init, Explore, Inspect, context assembly, and review reports remain CLI or internal derived workflows rather than becoming more MCP tools.
The local protocol cache contract is 300000/public for discovery and 0/private for the tool list; see MCP Compatibility for details.
Current maturity
Current package/source version is 0.39.1. Schema remains v12 and the public MCP surface remains exactly 17 tools. There is no automatic learning and no MCP tool #18. project init is the preferred first-project path. Default Explore is a Human-first view over existing repository-derived context and governed project knowledge. Current evidence and non-claims live in Production Status; published artifacts live in GitHub Releases.
Contributing
Read CONTRIBUTING.md for development and public-surface expectations, and SECURITY.md for vulnerability reporting.
Licensed under MIT.
agent-memory-bridgepip install agent-memory-bridge