Evidence before effect.
Auditable memory and exact-plan guarded actions for stateful AI agents.

Use cases · Quick start · macOS desktop · Flagship demo · Grok memory game · Scientific report · Guarded actions · Roadmap

A local-first Python engine for provenance-aware agent memory and optional guarded actions. The reference store is one SQLite file. Its security claims are deterministic and testable, but deliberately narrower than “the database is trusted” or “every external action is reversible.”

Think of AetnaMem as a governed continuity layer for AI systems. It helps an agent remember useful context, helps people understand where that context came from, and can carry evidence forward into reviewed decisions and approved changes. You can adopt one part without adopting the rest.

Most agent-memory integrations are discussed primarily as chat history, retrieval, or prompt caching. AetnaMem includes those practical concerns, but its distinctive value is the chain around them: source provenance, trust and quarantine, correction, deletion receipts, bounded context, independent audit verification, human approval, and an optional evidence-to-approved-change workflow. The same record can help an agent answer a question without being mistaken for authority to take an action.

That combination is useful when “the model remembered something” is not a sufficient operational explanation. It gives engineering teams a small local starting point, while giving managers, reviewers, and regulated organizations a path to inspect what informed a result and what was actually authorized. The measured OpenClaw/DeepSeek experiment also shows the practical side: cache-aware AetnaMem used 13.32% fewer prompt tokens and had 2.97% lower provider-reported cost than native MEMORY.md, with 20/20 answers in both arms. This is one controlled benchmark, not a universal saving; the raw trials and limitations are published for review.

The memory engine is usable today. The collaborative decision and EtD SDKs are experimental and need a real organizational pilot and external methodology review before clinical or regulatory claims. Multi-user products must also supply authentication, authorization, UI, networking, and deployment controls. See guarantee boundaries for the exact division of responsibility.

A local assistant desktop app: governed chat backed by a lightweight local Ollama model, visible/searchable memory, approval-gated file writes in a safe workspace, an in-browser file viewer/editor, and a live-verified audit chain. One launcher does everything — installs Ollama if missing, pulls the qwen3:1.7b model on first run, starts the service, and opens the dashboard in your browser already signed in (tokens ride in the URL fragment and never leave the browser).

git clone https://github.com/aetna000/aetnamem.git && cd aetnamem
chmod +x scripts/macos/aetnamem-desktop.command
open scripts/macos/aetnamem-desktop.command

macOS additionally seals the memory database encrypted at rest on quit, keyed through the Keychain. Details: aetnamem Desktop for macOS.

git clone https://github.com/aetna000/aetnamem.git && cd aetnamem
chmod +x scripts/linux/aetnamem-desktop.sh
./scripts/linux/aetnamem-desktop.sh

Needs python3 ≥ 3.10 (sudo apt-get install python3 / sudo dnf install python3) and curl. Ollama is installed via its official installer if missing. The database lives at ~/.local/share/aetnamem/memories.db (no at-rest sealing on Linux yet).

Double-click scripts\windows\aetnamem-desktop.bat, or run:

powershell -ExecutionPolicy Bypass -File scripts\windows\aetnamem-desktop.ps1

Needs Python 3.10+ (winget install Python.Python.3.12). Ollama is installed via winget if missing. The database lives at %LOCALAPPDATA%\aetnamem (no at-rest sealing on Windows yet).

1. Chat — the assistant runs fully on your machine with the local model. Tell it things worth remembering; ask it to draft files.
2. Approvals tab — when the assistant wants to write a file, the action is staged, never executed. Review the plan and click Approve & run or Deny.
3. Files tab — everything in your workspace (~/Aetnamem Workspace on macOS/Windows, ~/aetnamem-workspace on Linux). Markdown renders in the browser; any text file can be edited and saved in place.
4. Memory tab — every remembered fact with status, provenance, and trust tier; search and filter, including quarantined and forgotten records.
5. Settings — switch provider (local Ollama, OpenAI, DeepSeek, or any OpenAI-compatible endpoint), run the system check, and review the Data & security panel showing exactly where your data lives.

The desktop launchers use platform-native data locations: macOS stores a Keychain-protected sealed database under ~/Library/Application Support, Linux uses $XDG_DATA_HOME when set and otherwise ~/.local/share, and Windows uses %LOCALAPPDATA%. Workspace files are stored separately. Quit the service before copying SQLite data, and on macOS back up the encrypted database, its HMAC sidecar, and a recoverable copy of the Keychain key. See Data storage, backup, and restore for exact paths and restore instructions.

• Provenance is required. Extracted records link to their source episode; derived proposals instead cite existing episode or record IDs. Records also carry source, session, turn, time, confidence, status, and trust metadata.
• Classified untrusted content is quarantined. Records classified as webpage or tool output land quarantined until promote(). Correct origin classification is a host responsibility: an untrusted caller that may lie about source_type is outside this local API's trust boundary. promote() records a trust transition but does not authenticate human confirmation; protect or withhold that capability when the agent itself is untrusted.
• Recognized corrections supersede. When extraction assigns the same fact_key, the new trusted record supersedes the previous active record; the old record remains inspectable. Unrecognized or unkeyed contradictions are not automatically resolved.
• Memory content is logically purged. forget() tombstones matching records, clears their content and fact key, clears matching source episode text, removes FTS entries, and returns a deletion receipt. SQLite free pages, WAL files, filesystem snapshots, replicas, and backups require their own secure-erasure and retention controls.
• The audit log is independently checkable. Engine-generated memory and guarded-action transitions join a per-subject SHA-256 chain specified in audit-log specification. The standard-library independent verifier imports no aetnamem code. Hash chaining detects edits relative to a trusted head; externally anchored checkpoints are required to detect suffix deletion or replacement of the entire database.
• Sensitive values are separated on guarded paths. Core memory and guarded-action events use content digests and structural metadata. Raw action arguments and before-images live in an erasable payload table. retain_query_text=True stores raw recall queries, and the low-level log_action() method accepts caller-provided payloads, so callers must not place secrets or raw content there.

pip install aetnamem
# or, from a checkout:
pip install -e .

The package installs two console commands:

• aetnamem provides the memory CLI, the MCP server through aetnamem mcp, and guarded actions through aetnamem actions ….
• aetnamem-service starts the local assistant dashboard. Run aetnamem-service --help for workspace, database, network, and browser options. The cross-platform desktop launchers above additionally bootstrap Ollama and the recommended lightweight model.

aetna000 is the organization namespace only; it is not installed as a product command.

from aetnamem import Memory

m = Memory("./memories.db")          # or ":memory:"

m.remember("user-1", "My preferred airport is SFO.", session_id="s1")
m.remember("user-1", "Actually, use OAK as my preferred airport going forward.",
           session_id="s2")

m.recall("user-1", "Which airport should I fly from?")
# -> [{'content': "User's preferred airport is OAK.", 'status': 'active', ...}]

# Optional graph recall adds bounded multi-hop retrieval while preserving
# direct record fallback and the same governance rules.
m.recall("user-1", "What airport does my boss prefer?", use_graph=True)

m.forget("user-1", utterance="Forget my preferred airport.")
m.inspect("user-1")                  # full evidence dump, incl. audit chain check

The core verbs — remember, recall, list, forget, inspect, audit — plus promote (quarantine release), log_action (agent audit events), consolidate, persona, scenes, propose, checkpoint, and verify are available from Python and the CLI, so any process that can run a shell command is a client:

aetnamem remember   ./memories.db user-1 "My preferred airport is SFO." --session s1
aetnamem recall     ./memories.db user-1 "Which airport should I book from?"
aetnamem forget     ./memories.db user-1 --utterance "Forget my preferred airport."
aetnamem list       ./memories.db user-1 --all
aetnamem promote    ./memories.db user-1 rec_...
aetnamem log-action ./memories.db user-1 tool_call --payload '{"tool":"calendar"}'
aetnamem consolidate ./memories.db user-1
aetnamem persona    ./memories.db user-1
aetnamem scenes     ./memories.db user-1
aetnamem inspect    ./memories.db user-1
aetnamem audit      ./memories.db user-1
aetnamem checkpoint ./memories.db ./checkpoints.jsonl   # anchor this file externally
aetnamem verify     ./memories.db --checkpoints ./checkpoints.jsonl
python tools/verify_audit.py ./memories.db --checkpoints ./checkpoints.jsonl  # no aetnamem import

The standalone tools/verify_*.py commands are included in Git checkouts and source distributions. Wheel-only installs should use aetnamem verify and aetnamem actions verify, which cover the same integrity rules.

Guarded-actions mode turns a proposed tool mutation into a canonical hash-bound WorldPatch: exact operation digests, resource preconditions, adapter fingerprints, causal evidence, authority, approval, execution attempts, verification, compensation, and a receipt all share the subject's audit chain. Evidence that merely informed_by an operation is distinct from the host-attested authorized_by evidence that permits it.

The first reference adapter performs root-confined UTF-8 file writes and deletes. It is classified as verified compensatable, not transactionally reversible: aetnamem rechecks the before-state, executes only an exact approved plan, observes the after-state, and verifies any compensation against the captured before-state.

mkdir -p ./workspace

# Agent/host staging boundary: no reviewer key is needed here.
aetnamem actions stage ./memories.db user-1 filesystem write_text \
  --root ./workspace \
  --args '{"path":"report.md","content":"reviewed content"}' \
  --actor researcher-agent \
  --authority-id task-42 \
  --authority-digest 0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef

aetnamem actions show    ./memories.db act_...

# Separate trusted reviewer/executor shell. Persist this key securely.
export AETNAMEM_APPROVAL_KEY="$(python -c 'import secrets; print(secrets.token_hex(32))')"
aetnamem actions approve ./memories.db act_... --approver-label user-1
aetnamem actions commit  ./memories.db act_... --root ./workspace
aetnamem actions verify  ./memories.db act_...
python tools/verify_actions.py ./memories.db act_...  # no aetnamem import

Changing the persisted plan, adapter manifest, approval binding, expiry, or guarded file precondition prevents execution. Raw arguments, before-images, and provider receipts live in the erasable action payload plane; audit events contain only structural metadata and digests. Erase those payloads after their retention period with aetnamem actions purge-payloads ./memories.db act_.... If a process dies across an external execution boundary, use aetnamem actions recover ./memories.db act_...; it fences in-flight effects as uncertain and emits a recovery_required receipt instead of retrying.

Compatible external transaction journals can join the same forensic timeline without copying their raw arguments, snapshots, results, claimed actors, or client IDs into the audit plane:

aetnamem actions import-journal ./memories.db user-1 ./source-journal.db \
  --source-id production-agent

Imports are idempotent per source/transaction and are explicitly labeled unverified_operational_journal: importing external evidence does not upgrade its mutable status rows or claimed identities into aetnamem proof.

The HMAC approval key belongs in the human/reviewer process, not the agent-facing process. The --approver-label value is attribution; the shared key authenticates key possession, not that label. Likewise, CLI --authority-id/--authority-digest flags are only host-attested when a trusted host controls the staging command. The filesystem CLI is a reference vertical slice; the filter-only MCP gate is implemented, while automatic staging of arbitrary upstream MCP writes, Telegram review, additional execution providers, Firestore, and X adapters are tracked explicitly in the roadmap. Protocol and security details are in guarded-actions guide.

The optional aetnamem.decisions Python SDK adds namespace-scoped, multi-principal decision cases, immutable evidence links, conflict/recusal handling, ballots, deterministic outcomes, recommendation adoption, institutional approvals, and scoped change authorization. aetnamem.etd provides versioned clinical and generic Evidence-to-Decision templates plus deterministic reporting.

from aetnamem.decisions import ActorContext, DecisionEngine
from aetnamem.etd import clinical_etd_template

engine = DecisionEngine("organization.db")
chair = ActorContext("hospital-7", "principal-42")  # host-authenticated values
case = engine.create_case(
    chair,
    title="Medication reconciliation",
    template=clinical_etd_template(),
    content={"question": "Should we introduce pharmacist reconciliation?"},
    idempotency_key="request-001",
)

The host supplies authentication, users, UI, HTTP/TLS, notifications, and organization-level authorization. The SDK enforces case membership, recusal, idempotency, concurrency, exact revision binding, and audit transitions. The standard installation includes both the SQLite and PostgreSQL backends plus Ed25519 identity/receipt verification; the provider-neutral KMS adapter accepts a host-supplied AWS KMS client. Decision and COI payloads have independently configurable logical retention and signed purge receipts. These features do not claim GRADE, clinical, or regulatory compliance. Run aetnamem-etd-playground against SQLite or PostgreSQL and verify its export with aetnamem-etd-verify.

See the decision workflow specification, EtD profile, and host integration guide. Organizations preparing real users should also use the pilot and methodology-review runbook.

aetnamem mcp currently serves memory verbs only as MCP tools over stdio: newline-delimited JSON-RPC implemented with the standard library only and does not require an additional MCP framework. Defaults: database at ~/.aetnamem/memories.db (override with --db or $AETNAMEM_DB) and subject default (--subject), so single-user personal agents need no per-call subject wiring. It is not yet an action interception gateway and does not prevent an agent from calling other write tools directly.

Claude Code:

claude mcp add aetnamem -- aetnamem mcp

Claude Desktop / any host with JSON MCP config (OpenClaw's MCP bridge takes the same command + args shape):

{
  "mcpServers": {
    "aetnamem": {
      "command": "aetnamem",
      "args": ["mcp", "--db", "/home/you/.aetnamem/memories.db"]
    }
  }
}

The agent gets memory_remember, memory_recall, memory_recall_block (bounded prompt-injection block), memory_persona, memory_context_pack (host-neutral stable/dynamic context), memory_capture (auto-capture with digest-only assistant/tool logging), memory_list, memory_forget, memory_promote, memory_audit, memory_verify, and memory_graph_status, memory_graph_merges, memory_graph_history, and memory_log_action. Graph merge decisions are deliberately absent from MCP; they require the reviewer-authenticated dashboard/HTTP surface or explicit CLI use.

subject_id is a storage scope chosen by the caller, not an authenticated tenant identity. Likewise, exposing memory_promote lets the agent request a promotion; use a trusted approval layer or omit that tool when promotion must be human-only.

Grok/xAI users: the Grok/xAI guide shows how to expose aetnamem as xAI function-calling tools today, with a local playground that lets Grok search, capture, forget, and audit memory while the engine keeps provenance and deletion receipts. xAI Remote MCP is the deployment path once the local stdio MCP server is exposed behind an HTTP/SSE gateway.

OpenClaw users: the native integration is a native plugin that adds automatic memory — auto-recall injection before every prompt and auto-capture after every turn — on top of the same engine and audit chain. The policy gates run server-side, so a hostile webpage summarized by the agent still cannot plant durable memory, deletion still returns receipts, and you can independently audit the same SQLite file with aetnamem verify or tools/verify_audit.py while the agent uses it. Full tool catalog, host configs, and troubleshooting: integration guide.

Install the native integration in three commands:

python3 -m pip install --upgrade aetnamem
openclaw plugins install npm:openclaw-memory-aetnamem@latest --pin
openclaw aetnamem setup --single-user --subject you

The setup applies conservative context budgets and enables the required conversation hooks. It does not rewrite MEMORY.md; verify recall before removing duplicated native memory. One plugin instance currently has one fixed subject and must not be shared between authenticated users.

By default, each prompt receives no more than three matching memories within a 1,200-character recall block plus a 600-character persona block, and unrelated queries receive no recall block.

In the checked-in three-arm OpenClaw 2026.7.1-2 + DeepSeek V4 Flash follow-up, 20 fresh-session tasks per arm used 596,581 prompt tokens with a 19,489-character native MEMORY.md, 521,858 with the current AetnaMem layout, and 517,118 with cache-aware AetnaMem. The optimized arm therefore used 79,463 fewer prompt tokens (13.320%) and cost 2.968% less than native, with 20/20 correct answers in every arm and 20/20 target retrieval in both AetnaMem arms. It did not increase absolute cache reads over current AetnaMem; its additional 0.908% token and 1.190% cost improvement came from the optimized bundle's smaller model-visible surface. An earlier run produced a 0.674% cost increase under a different cache mix, so neither bill is a universal claim. The raw trials, protocol, limitations, and reproduction command are checked in; the plugin README also includes the two-minute recall demo.

The rule is: MCP first, native adapter only when it adds lifecycle hooks. Do not fork the memory semantics per host. aetnamem should stay the auditable engine; framework integrations should be thin wrappers that call the same MCP/Python verbs and preserve the same audit trail.

For any MCP-capable host, start with:

aetnamem mcp --db ~/.aetnamem/memories.db --subject you

Then configure the host to expose the memory_* tools. This is the right first path for Hermes, Claude Desktop, Claude Code, and any framework that can launch a stdio MCP server. The dedicated Hermes guide includes the exact commands and explains the boundary between tool-based memory and automatic prompt injection.

For an automatic adapter, call one provider-neutral operation before the model invocation:

pack = memory.build_context_pack("user-42", current_query)
system_prefix += "\n" + pack["stable_context"]
current_turn += "\n" + pack["dynamic_context"]

The same contract is available as CLI aetnamem context-pack and MCP tool memory_context_pack. The stable block is deterministic and suitable for a stable system-prefix position; the dynamic block is selective and belongs near the current turn. Records already present in the stable block are not repeated in the dynamic block. This layout can preserve provider prefix-cache reuse while keeping the total prompt bounded, but it cannot guarantee cache hits: cache boundaries, minimum cacheable lengths, TTLs, and billing remain provider and host behavior.

Build a native adapter only when the framework gives useful hooks:

Native adapters should pass the host's session_id and turn_id whenever available, so memory reads, writes, tool calls, forgets, and user-visible responses line up in one audit timeline.

Priority targets:

The adapter directory should stay organized by host:

integrations/
  openclaw/
  hermes/
  langgraph/
  openai-agents/
  crewai/
  microsoft-agent-framework/
  llamaindex/
  haystack/

Each adapter should document the same guarantees: untrusted content stays quarantined, deletion returns receipts, recall injection is bounded and audited, and the SQLite database can still be verified externally with aetnamem verify or tools/verify_audit.py.

The architecture separates mutable content tables from an engine-append-only, hash-chained audit table. This can support accountability, access, rectification, and deletion workflows, but it is not compliance certification or legal advice. forget() performs a logical purge in the live database; it does not by itself sanitize SQLite free pages, WAL files, backups, exports, or external replicas. Checkpoint placement, retention, secure erasure, access control, identity, lawful basis, and jurisdiction-specific requirements remain deployment responsibilities. See the audit-log specification for the precise integrity threat model.

• L0 — episodes: raw turns, append-only, purged by deletion.
• L1 — records: extracted facts with provenance.
• L2 — scenes: deterministic per-session view (aetnamem scenes).
• L3 — persona: live-derived snapshot of active facts (aetnamem persona, MCP memory_persona) — no cached persona is stored; each generated snapshot carries its source record IDs.
• Derived proposals: external LLM/batch jobs submit candidates via aetnamem propose / Memory.propose_facts(); they land quarantined with mandatory evidence links and only activate through promote().

Recall has top-k semantics. SQLite FTS5 selects a bounded candidate set (200 by default), trust and recency priors rank it, and the best limit records are returned. Quarantined, superseded, and tombstoned records are never candidates. Retrieval events record the algorithm, cap, candidate count, score inputs, and a bounded ledger (the first 50 plus every returned result). A versioned digest binds the selected retrieval fields to the audit chain. Pass min_score= to drop weak matches.

Graph recall is opt-in with use_graph=True, CLI --graph, or service AETNAMEM_GRAPH_RECALL=1. It extracts a conservative entity/edge index from governed records, seeds it with FTS5, spreads through at most two bounded hops, and blends the result with direct record recall. Returned graph hits include their path evidence. The index inherits quarantine, promotion, supersession, and forgetting; it can be recreated at any time:

aetnamem graph-backfill ./memories.db user-1
aetnamem recall ./memories.db user-1 "What airport does my boss prefer?" --graph
aetnamem graph-consolidate ./memories.db user-1
aetnamem graph-merges ./memories.db user-1 --status pending
aetnamem graph-inspect ./memories.db user-1

v0 extraction is deterministic (generic sentence patterns: "my X is Y", "use Y as my X", "remember that …", "I avoid …") so that policy failures are debuggable, not probabilistic. The local Python API, CLI, MCP server, deterministic consolidation, persona snapshots, scenes, checkpoints, and independent memory verifier are implemented. An optional, deterministic graph index provides bounded multi-hop recall with path evidence and direct-record fallback. Guarded Actions additionally ships an action ledger, exact-plan shared-key approvals, filesystem reference adapter, recovery fencing, external journal import, and independent action verifier. A fail-closed, filter-only MCP gate is implemented; automatic conversion of arbitrary upstream writes into staged actions, authenticated host identity, encrypted payloads, LLM-backed graph extraction, public HTTP deployments, and additional storage backends remain roadmap work — see the roadmap. The policy gates in aetnamem/core/policy.py are the product; nothing in the engine may reference the vocabulary of a benchmark scenario.

• 0.4.1 release notes — provider-neutral cache-aware context packs, OpenClaw npm release, Hermes integration, measured token/cost results, and claims boundaries.
• 0.4.0 release notes — evidence-to-approved-change capabilities, installation, validation, and claims boundary.
• macOS desktop guide — local dashboard, onboarding checks, provider setup, approval UI, safe workspace, Keychain secrets, and encrypted at-rest database sealing.
• Data storage and backup — default database/workspace paths on macOS, Linux, and Windows, plus safe backup, key recovery, restore steps, and encryption boundaries.
• Integration guide — full CLI reference (every command, flags, output shapes, exit codes) and MCP server reference (transport, flags, tool catalog, host configs for Claude Code / Claude Desktop / OpenClaw-style bridges, security properties, troubleshooting).
• OpenClaw setup — visual (Mermaid) walkthrough of wiring aetnamem into OpenClaw or any MCP host: setup flow, runtime sequence, the quarantine gate, and the external audit loop.
• Hermes Agent guide — MCP setup, tool-based memory, automatic context-pack integration, caching expectations, and multi-user boundaries.
• Grok/xAI guide — Grok/xAI function-calling quickstart, local playground, and Remote MCP deployment notes.
• Graph memory design — implemented Phase 0–4 graph index: entities and typed edges over governed records, bounded seed+spread recall, reviewer-gated reversible merges, scheduled consolidation, cold history partitions, and incremental audit verification.
• Auditing guide — how to use the auditability: checkpoint cadence and anchoring recipes, verifying after an incident, handling erasure/access/rectification requests with receipts, reviewing quarantine, logging agent actions onto the same chain, and what to hand an external auditor.
• Audit-log specification — the frozen wire format: canonical serialization, hash preimages, chain/checkpoint/receipt verification rules, and the threat-model table.
• Guarded actions — action modes, authority boundaries, state transitions, guarantees, and non-guarantees.
• Collaborative decision workflow — generic cases, revisions, evidence lineage, ballots, adoption, authorization, concurrency, receipts, and trust boundaries.
• Evidence-to-Decision profile — versioned EtD criteria, artifact chain, report surface, and methodology boundary.
• Decision host integration — authenticated-host contract, namespace derivation, SQLite/PostgreSQL deployment, signed identity, retention, and approved-change bridge.
• EtD pilot and methodology review — production entry criteria, multi-user test protocol, acceptance evidence, and independent-review package.
• Channels and governed outbound proposal — provider-neutral intake, optional business review, and evidence-bound external adapters; proposed, not implemented.
• Inference engineering memory proposal — provider-neutral local and Hugging Face inference, governed run evidence, strict comparison, engineering decisions, deployment receipts, and incident reconstruction; proposed, not implemented.
• Roadmap — completed foundation work and remaining product, provider, security, and interface tasks.
• Architecture plan — architecture plan and roadmap.

Development is gated on MemoryStackBench's seven_sins_v0_1 suite (webpage poisoning, retention after deletion, missing provenance, stale temporal updates, overgeneralization). Current score: 33/33, with unit tests covering the same gates on non-benchmark vocabulary to keep the score honest.

git clone https://github.com/aetna000/MemoryStackBench.git
cd MemoryStackBench
cp /path/to/aetnamem/bench/adapters/aetnamem.py memorybench/adapters/aetnamem.py
cp /path/to/aetnamem/bench/targets/aetnamem.yaml targets/aetnamem.yaml
PYTHONPATH=/path/to/aetnamem:$PWD \
python -m memorybench.cli run \
  --target targets/aetnamem.yaml \
  --suite suites/seven_sins_v0_1 \
  --out runs/aetnamem-local

AGPL-3.0 (see the license). Anyone may use aetnamem, including commercially, but derivative works — including software that serves aetnamem over a network — must be released under the same terms.