Skip to content

README.md

Latest commit

 

History

History
330 lines (234 loc) · 23.8 KB

README.md

File metadata and controls

330 lines (234 loc) · 23.8 KB

agent-memory

CI npm PyPI Python License: MIT

Local-first memory for AI agents.

agent-memory gives Hermes, Codex-style CLIs, Claude-style CLIs, and custom agent harnesses a shared SQLite memory runtime with curation, provenance, retrieval, prompt rendering, and regression evaluation.

It is intentionally small and local-first: your memory database lives on your machine unless you choose to sync or copy it elsewhere.

Why use it?

Most agent memory systems end up as raw logs, ad-hoc notes, or one-shot RAG. agent-memory separates durable knowledge into semantic facts, procedures, episodes, source records, scopes, and lifecycle states so an agent can remember useful context without blindly stuffing every transcript into future prompts.

Use it when you want:

  • one user-level memory store shared across multiple agent harnesses
  • local SQLite storage instead of a hosted memory service
  • approved-only prompt context by default
  • candidate/disputed/deprecated memory review flows
  • source/provenance metadata for every curated memory
  • bounded prompt rendering for Hermes/Codex/Claude wrappers
  • retrieval regression fixtures with lexical/source baselines and failure triage

Docs for first-run and operational validation:

30-second install

Recommended path for CLI agent users:

npm install -g @cafitac/agent-memory
agent-memory bootstrap
agent-memory doctor

What this does:

  • installs the agent-memory command
  • initializes ~/.agent-memory/memory.db when missing
  • creates or merges the Hermes hook config at ~/.hermes/config.yaml
  • preserves existing Hermes hooks and appends the agent-memory pre-LLM hook
  • lets you verify setup with agent-memory doctor

Python-first alternatives:

pipx install cafitac-agent-memory
agent-memory bootstrap
agent-memory doctor
uv tool install cafitac-agent-memory
agent-memory bootstrap
agent-memory doctor

First memory in 60 seconds

DB=~/.agent-memory/memory.db

agent-memory init "$DB"
agent-memory create-fact "$DB" "agent-memory" "primary-install-path" "npm install -g @cafitac/agent-memory" "user:default"
agent-memory approve-fact "$DB" 1
agent-memory retrieve "$DB" "How should I install agent-memory?" --preferred-scope user:default

Normal retrieval is approved-only by default. Candidate, disputed, and deprecated memories stay out of prompt context unless you intentionally ask for a forensic view:

agent-memory retrieve "$DB" "What obsolete install notes exist?" --status all
agent-memory review conflicts fact "$DB" "agent-memory" "primary-install-path"

Review transitions can carry operator context and remain inspectable later. If one fact replaces another, record the replacement chain so stale facts can be explained without entering normal retrieval:

agent-memory review approve fact "$DB" 1 --reason "Verified from current setup guide." --actor maintainer --evidence-ids-json '[1]'
agent-memory review history fact "$DB" 1
agent-memory review supersede fact "$DB" 1 2 --reason "Newer source replaces the old install path." --actor maintainer --evidence-ids-json '[2]'
agent-memory review replacements fact "$DB" 1
agent-memory review explain fact "$DB" 1

review explain combines the current status, default retrieval visibility, transition history, same claim-slot alternatives, and replacement chain into one decision context so a reviewer can see why a stale or conflicting fact is hidden.

For read-only relation graph inspection, use graph inspect. This is an operator/debug view over stored Relation edges; it does not change retrieval behavior or mutate memory state:

agent-memory graph inspect "$DB" fact:1 --depth 1
agent-memory graph inspect "$DB" fact:1 --depth 2 --limit 50

The JSON output includes the start ref, visited node refs, relation edges, traversal depth per edge, and a read_only: true marker. It is intended as a safe graph-foundation slice before enabling any broader graph traversal in default retrieval.

For local dogfood and noise monitoring, retrievals can leave a secret-safe observation log. Normal retrieve only records an observation when explicitly asked; the Hermes pre-LLM hook records one automatically in the local SQLite DB for real turns. Observations store a query hash, selected memory refs, top memory ref, response mode, scope, and surface. They do not store the raw query text or a query preview. Deterministic hermes hooks doctor/test pre-LLM payloads exercise context injection but are skipped as dogfood observations so synthetic weather prompts do not pollute the audit.

agent-memory retrieve "$DB" "How should I install agent-memory?" --preferred-scope user:default --observe cli
agent-memory observations list "$DB" --limit 20
agent-memory observations audit "$DB" --limit 200 --top 10 --frequent-threshold 3
agent-memory observations empty-diagnostics "$DB" --limit 200 --top 10 --high-empty-threshold 0.5
agent-memory observations review-candidates "$DB" --limit 200 --top 10 --frequent-threshold 3
agent-memory activations summary "$DB" --limit 200 --top 20 --frequent-threshold 3
agent-memory activations reinforcement-report "$DB" --limit 200 --top 20 --frequent-threshold 3
agent-memory activations decay-risk-report "$DB" --limit 200 --top 20 --frequent-threshold 3
agent-memory consolidation candidates "$DB" --limit 200 --top 20 --min-evidence 2
agent-memory consolidation explain "$DB" <candidate-id> --limit 200 --min-evidence 2
agent-memory consolidation promote fact "$DB" <candidate-id> \
  --subject-ref "agent-memory" \
  --predicate "prefers" \
  --object-ref-or-value "explicit human-reviewed promotion" \
  --scope project:agent-memory
agent-memory consolidation promotions report "$DB" --limit 50
agent-memory dogfood baseline "$DB" --output-json
agent-memory traces record "$DB" --surface cli --event-kind user_correction --summary "sanitized trace summary" --scope project:agent-memory
agent-memory traces list "$DB" --surface cli --limit 20
agent-memory traces retention-report "$DB" --max-trace-count 10000

Use the observation log and audit report to spot frequently injected or surprising memories before changing retrieval behavior. The audit output is read-only JSON with surface/scope counts, empty-retrieval count and ratio, quality warnings such as low_observation_count or high_empty_retrieval_ratio, top injected memory refs, current status for known refs, per-ref observation windows, and simple signals such as frequently_injected and current_status_not_approved. observations empty-diagnostics is read-only and focuses specifically on empty retrievals: it groups empty-heavy observations by surface, preferred scope, and status filter with segment ratios, sample observation ids, observation windows, and next-step hints for checking scope mismatches or missing approved memory coverage before changing rankers. observations review-candidates is also read-only; it turns the top audit refs into forensic candidates with top-level observation_count/candidate_count, fact review explanations, status-history summaries, replacement-chain hints, graph-neighborhood summaries, and copy-paste follow-up commands such as review explain, review replacements, and graph inspect.

The next consolidation layer is an experimental trace and activation substrate, not automatic memory creation. experience_traces stores low-cost local event traces behind explicit write APIs, the manual traces record CLI, and an explicit Hermes hook opt-in flag. Hermes trace recording remains disabled by default; enable it only by adding --record-trace to the hermes-pre-llm-hook command or rendering/installing a hook snippet with --record-trace. Hook trace rows store a content hash, hashed session ref, safe adapter metadata such as platform/model, related retrieved memory refs, and retention metadata. They intentionally do not store raw prompts/transcripts/user messages and do not change approved-memory retrieval or ranking. Synthetic hermes hooks doctor/test payloads are skipped, and trace write failures are non-blocking. traces list is read-only JSON and supports --surface, --event-kind, and --scope filters for local dogfood. traces retention-report is also read-only; it summarizes retention-policy counts, expired trace refs, expirable traces missing expires_at, and volume warnings without deleting traces, promoting memories, or printing trace metadata/summary text.

Stage C starts with memory_activations, a local-only internal substrate that distinguishes "a trace happened" from "a memory was retrieved/activated." Retrieval observations now bridge into activation events: selected memory refs create retrieved activations, while empty retrievals create empty_retrieval negative evidence. Activation rows store refs, observation links, scope, strength, and sanitized metadata only; they do not store raw queries or prompt previews, and they do not change retrieval ranking or long-term memory status.

agent-memory activations summary "$DB" --limit 200 --top 20 --frequent-threshold 3 is the first read-only Stage C reporting surface. It summarizes activation counts, activation windows, surfaces/scopes, status counts for top refs, empty-retrieval negative evidence, and top memory refs with advisory signals such as frequently_activated, likely_reinforcement_candidate, or current_status_not_approved. It remains local-only and read-only: no raw queries, no prompt previews, no ranker changes, no memory status mutation, and no automatic long-term promotion.

agent-memory activations reinforcement-report "$DB" --limit 200 --top 20 --frequent-threshold 3 adds a deterministic read-only score over activation refs. The report explains every candidate score with factor breakdowns for repetition, strength, status trust, surface/scope diversity, and graph connectivity, plus penalties for deprecated/disputed/missing refs and supersession/replacement relations. It is advisory only: it does not change retrieval ranking, memory status, or long-term promotion state.

agent-memory activations decay-risk-report "$DB" --limit 200 --top 20 --frequent-threshold 3 adds the paired read-only decay-risk view. It flags weak/low-use/low-connectivity/stale activation refs with factor breakdowns, but protects approved, frequently activated, connected refs from naive age-only recommendations. The output suggests review/explanation commands only; it does not delete traces, deprecate memories, alter status, or change retrieval ranking.

agent-memory consolidation candidates "$DB" --limit 200 --top 20 --min-evidence 2 starts Stage D as a read-only trace clustering diagnostic. It groups sanitized experience_traces with deterministic scope/memory/summary keys, emits stable candidate fingerprints, evidence windows, surfaces/scopes, safe summaries, related memory/observation refs, current status and activation reinforcement context, guessed memory type, and risk flags. It does not create, approve, reject, snooze, or mutate memories, and it never prints raw prompts, queries, transcripts, or query previews.

agent-memory consolidation explain "$DB" <candidate-id> --limit 200 --min-evidence 2 expands one candidate into an auditable read-only review packet. The explanation repeats the stable candidate payload, shows why the traces were grouped, exposes safe trace ids/windows/summaries, supporting activation/status signals, guessed memory type rationale, risk flags, and explicit review-state guardrails. Unknown candidate ids return JSON with found: false and a non-zero exit. The command is still advisory-only: it does not promote, approve, reject, snooze, or mutate memories, and it never prints raw trace payloads, prompts, transcripts, or query_preview values.

agent-memory consolidation promote fact "$DB" <candidate-id> --subject-ref ... --predicate ... --object-ref-or-value ... --scope ... starts Stage E with an explicit human-reviewed promotion action for semantic facts. The reviewer must supply the final fact fields; the command uses the candidate only as safe provenance, creating a local consolidation_candidate source from candidate id, trace ids, related observation ids, and safe summaries. By default it creates a candidate fact, so default retrieval remains approved-only; pass --approve --actor ... --reason ... only after explicit review to approve the new fact and log the status transition. It also records graph lineage edges from the candidate fingerprint to the promoted fact (promoted_to) and from the fact to its generated provenance source (has_promotion_provenance) so graph inspect "$DB" <candidate-id> --depth 2 can explain how a durable memory came from reviewed consolidation evidence. Unknown candidate ids return JSON with promoted: false and do not create sources, facts, or lineage relations. Procedure/preference promotion and conflict preflight remain future Stage E slices.

agent-memory consolidation promotions report "$DB" --limit 50 adds the first read-only audit surface over those manual promotions. It lists promoted semantic facts, candidate fingerprints, generated provenance source ids, safe summaries/trace ids/observation ids, status counts, approval history, and the expected graph lineage relation refs without mutating facts, sources, relations, status transitions, retrieval ranking, or trace state. It is intended for review and release dogfood after promotion; it never prints raw prompts, transcripts, query previews, or raw trace metadata.

agent-memory dogfood baseline "$DB" --output-json composes the same read-only observation reports with package version, database path/schema metadata, memory status counts, sanitized Hermes hook doctor metadata, a non-executed local E2E marker, and suggested next steps. The baseline intentionally omits raw queries, query previews, prompt text, full Hermes config, and the bootstrap command so outputs can be pasted side by side during later trace/consolidation PRs. Treat all of these reports as local operator telemetry, not a synced analytics feature or an automatic cleanup workflow.

Hermes quickstart

For most Hermes users:

npm install -g @cafitac/agent-memory
agent-memory bootstrap
agent-memory doctor
hermes hooks doctor

On first real Hermes use, Hermes may ask you to approve the shell hook or require --accept-hooks depending on your local Hermes policy.

The installed hook calls:

agent-memory hermes-pre-llm-hook ~/.agent-memory/memory.db --top-k 1 --max-prompt-lines 6 --max-prompt-chars 800 --max-prompt-tokens 200 --max-verification-steps 1 --max-alternatives 0 --max-guidelines 1 --no-reason-codes

The hook receives the Hermes event JSON on stdin, retrieves relevant approved memories, and returns bounded ephemeral context for the current prompt. It does not write back to Hermes session storage. agent-memory bootstrap uses the conservative Hermes preset by default: one top memory, small prompt budgets, no alternative-memory detail in the prompt, no reason-code noise, and fail-closed behavior if retrieval is unavailable.

If you only want to inspect the YAML snippet and not modify config:

agent-memory hermes-hook-config-snippet ~/.agent-memory/memory.db

If you want explicit paths and budgets:

agent-memory hermes-install-hook ~/.agent-memory/memory.db --config-path ~/.hermes/config.yaml --preset conservative --timeout 8
agent-memory hermes-install-hook ~/.agent-memory/memory.db --config-path ~/.hermes/config.yaml --preset balanced

Use --preset balanced if you intentionally want the older, more verbose hook shape (--top-k 3, larger budgets, and reason codes). Explicit flags such as --top-k, --max-prompt-tokens, or --no-reason-codes override the selected preset.

Codex and Claude prompt wrappers

For harnesses that want a plain prompt prefix rather than a Hermes hook response:

agent-memory codex-prompt ~/.agent-memory/memory.db "What should I remember about this project?" --preferred-scope user:default
agent-memory claude-prompt ~/.agent-memory/memory.db "What should I remember about this project?" --preferred-scope user:default

The command prints prompt text only, so wrappers can prepend it to the live user request before invoking Codex, Claude Code, or another CLI.

This repository also includes source-checkout helper scripts for maintainers:

python scripts/run_codex_with_memory.py ~/.agent-memory/memory.db "What should I do next?" --dry-run
python scripts/run_claude_with_memory.py ~/.agent-memory/memory.db "What should I do next?" --dry-run

End users should prefer the installed agent-memory command unless they are developing this repository.

Data and privacy model

  • Default database: ~/.agent-memory/memory.db
  • Default Hermes config path: ~/.hermes/config.yaml
  • Storage: local SQLite
  • Network behavior: the core CLI does not upload your memory database to an agent-memory cloud service
  • Prompt policy: approved memories are retrieved by default; candidate/disputed/deprecated memories are excluded unless requested
  • Scope policy: user:default is the recommended durable cross-project scope; Hermes can also derive privacy-preserving cwd:<hash> scopes without exposing raw local paths in prompt context

See PRIVACY.md and SECURITY.md for the external-user trust model, sensitive-data guidance, and vulnerability reporting instructions.

Uninstall and rollback

Uninstall the CLI:

npm uninstall -g @cafitac/agent-memory
# or
pipx uninstall cafitac-agent-memory
# or
uv tool uninstall cafitac-agent-memory

Remove the Hermes hook by editing ~/.hermes/config.yaml and deleting the agent-memory hermes-pre-llm-hook ... command from hooks.pre_llm_call.

Keep or remove local data explicitly:

# inspect first
ls -lh ~/.agent-memory/memory.db

# destructive: removes the local memory database
rm ~/.agent-memory/memory.db

agent-memory bootstrap backs up changed Hermes config files to *.agent-memory.bak when it modifies an existing config.

Retrieval evaluation and regression gates

agent-memory includes a fixture-based retrieval evaluator so retrieval behavior can be tested like application code:

agent-memory eval retrieval ~/.agent-memory/memory.db tests/fixtures/retrieval_eval
agent-memory eval retrieval ~/.agent-memory/memory.db tests/fixtures/retrieval_eval --baseline-mode lexical
agent-memory eval retrieval ~/.agent-memory/memory.db tests/fixtures/retrieval_eval --baseline-mode lexical --format text
agent-memory eval retrieval ~/.agent-memory/memory.db tests/fixtures/retrieval_eval --fail-on-regression

Supported baseline modes include:

  • lexical: preferred-scope lexical comparator
  • lexical-global: lexical comparator that ignores preferred scope
  • source-lexical: lexical comparator over linked source content within preferred scope
  • source-global: source-linked comparator that ignores preferred scope

Reports include per-task retrieved IDs, expected hits, missing IDs, avoid hits, pass/fail state, aggregate summaries, soft-threshold advisories, and failure triage details such as snippets, lifecycle status, scopes, and policy signals. Every JSON result also includes an advisory_report with severity, affected task IDs, and recommended next actions. Text reports render the same advisory report as terminal-friendly guidance for maintainers reviewing failed retrieval tasks; JSON is the stable machine-readable surface.

The evaluator calls the real retrieval path but suppresses retrieval bookkeeping side effects while it runs. Eval tasks do not increment retrieval_count, reinforcement_count, or last_accessed_at, so fixture order and repeated local/CI runs do not perturb later ranking results.

Current maturity

agent-memory is alpha software, but the public install path is validated.

What is ready:

  • npm and PyPI releases from the same versioned source
  • GitHub Release artifacts
  • CI and release metadata checks
  • published-install smoke checks
  • local SQLite storage
  • Hermes bootstrap/doctor flow
  • Codex/Claude prompt rendering commands
  • approved-only retrieval policy by default
  • retrieval regression fixtures and diagnostic reports

Known limitations:

  • no hosted sync service
  • no built-in encryption-at-rest wrapper around the SQLite file
  • no automatic secret detection/redaction before users create memories
  • no stable 1.0 API guarantee yet
  • advanced graph/semantic retrieval behavior is still evolving
  • multi-machine sharing is currently a user-managed file/sync concern

Development

git clone https://github.com/cafitac/agent-memory.git
cd agent-memory
uv run pytest tests/ -q
uv run python scripts/check_release_metadata.py
uv run python scripts/smoke_release_readiness.py
uv run pytest tests/test_published_install_smoke.py -q
npm pack --dry-run

Release automation expects protected main: feature PRs and release-sync PRs run ci.yml, while auto-release.yml ignores docs/workflow-only pushes and publish.yml runs only from version tags or explicit manual dispatch. To keep GitHub Actions minutes bounded, publish.yml performs release metadata validation, package build, npm dry-run, npm/PyPI publish, and GitHub Release creation, but it no longer repeats the full pytest suite or runs the slow real-registry install matrix by default.

If a release needs extra external-install validation, run the opt-in published-install-smoke workflow manually with gh workflow run published-install-smoke.yml -f version=<version>, or dispatch publish.yml with -f run_published_install_smoke=true. The smoke verifies the exact npm/PyPI version through npm registry lookup, npx, npm exec, uvx, and pipx. The smoke script treats early resolver/package-index misses as propagation-like failures and applies exponential backoff before failing; failure artifacts include npm/PyPI registry probe diagnostics so maintainers can tell whether metadata is visible while installers are still stale.

If the auto-release workflow cannot push its bumped metadata commit directly, it opens a release-sync/vX.Y.Z PR instead. After that PR is merged, the same workflow tags the synced version and dispatches fast publish.yml with run_published_install_smoke=false, keeping the release path automated without requiring a permanent branch-protection bypass. The fallback is safe to rerun: if the release-sync/vX.Y.Z branch or PR already exists, the workflow reuses it instead of failing on a non-fast-forward push or opening a duplicate PR. When it creates a new release-sync PR, it also dispatches ci.yml on that bot-created branch and comments with the validation handoff because GitHub can suppress automatic PR checks for bot-created refs.

Useful source-checkout commands:

uv run python -m agent_memory.api.cli --help
uv run python -m agent_memory.api.cli hermes-bootstrap /tmp/agent-memory.db --config-path /tmp/hermes-config.yaml
uv run python -m agent_memory.api.cli hermes-doctor /tmp/agent-memory.db --config-path /tmp/hermes-config.yaml

Repository docs

  • docs/install-smoke.md: published install smoke recipes
  • SECURITY.md: vulnerability reporting and local security model
  • PRIVACY.md: local data, prompt, and hook privacy model
  • CONTRIBUTING.md: contribution workflow
  • .dev/: AI-authored drafts, design spikes, research notes, and unapproved plans
  • docs/: human-reviewed promoted documentation

License

MIT. See LICENSE.