aetnamem

Install beside OpenClaw. Mirror first. Switch only after proof.
Searchable, auditable memory takeover with a frozen native snapshot and one-command rollback.

Start safely · Watch the OpenClaw demo · Agent skills · Safe Switch guide · Current status · Python / MCP · Research

AetnaMem is a local-first memory and evidence layer for agents. Version 0.7.0a2 is an experimental OpenClaw shadow-and-takeover preview: install it beside an existing OpenClaw agent, mirror the native Markdown memory without changing prompts, search and trace the governed copy, try bounded recall internally, and activate only after the mirror and rollback evidence verify.

The generic Python Memory API, aetnamem mcp, existing SQLite databases, and normal OpenClaw integration remain compatible. The matching OpenClaw plugin is 0.5.0-experimental.1. These are prereleases, not the final stable 0.6.1 / 0.4.0 packages.

Adopt AetnaMem without a blind switch

Your current OpenClaw memory remains authoritative during evaluation:

1. Install beside OpenClaw. AetnaMem installs and verifies its matching bridge; nothing is cut over.
2. Copy, then mirror native memory. AetnaMem first takes a verified, byte-for-byte baseline of every recognized persistent memory source: MEMORY.md, memory/, USER.md, pinned workspace instructions and workspace skills. Only then are searchable Markdown sources copied into an isolated AetnaMem mirror with file hashes and line provenance.
3. Shadow real prompts. AetnaMem computes bounded recall without injecting it or making an extra model call.
4. Inspect the copy. Search ordinary words, follow a trace, verify the audit chain, and review the exact source manifest in the dashboard.
5. Activate with a verified freeze. AetnaMem takes a second complete switch-time snapshot, verifies every byte digest, disables duplicate native memory retrieval and session-memory writing, and becomes the bounded supplemental-memory path without changing the agent-facing memory_search / memory_get names.
6. Roll back exactly. Restore native files, memory slot, session-memory hook and the prior plugin configuration without deleting trial evidence. Memories captured while AetnaMem was active are exported into a hash-verified native Markdown file. If anything recreated the native paths during takeover, those divergent files are preserved in the trial evidence directory before restoration.

Start the side-by-side OpenClaw trial:

# 1. Install the AetnaMem engine.
python -m pip install --pre aetnamem==0.7.0a2
aetnamem --version

# 2. Let AetnaMem install and verify the matching OpenClaw bridge.
aetnamem openclaw install

# The verified installer leaves OpenClaw authoritative and AetnaMem in shadow.
aetnamem trial status
aetnamem openclaw memory status
aetnamem openclaw memory search "TypeScript preference"
aetnamem openclaw memory trace "TypeScript preference"

aetnamem openclaw install records the exact engine executable, installs the matching npm bridge internally, restarts and probes the gateway, and starts a shadow trial. It synchronizes changed native memory on status, search, trace and live shadow recall. If installation verification fails, it restores the prior OpenClaw plugin configuration. Shadow mode does not change model context or make provider calls. Users do not run npm directly.

Run the friendly dashboard in the foreground or as a background process:

aetnamem dashboard

aetnamem dashboard daemon start --port 8766
aetnamem dashboard daemon open    # launch the protected URL in your system browser
aetnamem dashboard daemon status
aetnamem dashboard daemon restart
aetnamem dashboard daemon stop
aetnamem dashboard daemon remove  # service record only; memory is preserved

The daemon creates a high-entropy local access key and keeps that protected URL valid for the lifetime of the process. daemon open avoids editor link handlers and launches the operating system's browser directly. A restart rotates the key and invalidates earlier sessions. The local state and log files that contain access metadata are mode 0600; the server remains loopback-only and also requires an HttpOnly session cookie, CSRF token, and valid Origin for changes.

There are only two customer-visible states:

# OpenClaw memory is active while AetnaMem mirrors it.
aetnamem trial status

# After inspecting and searching the verified mirror:
aetnamem trial activate
aetnamem openclaw memory status

# Make the frozen OpenClaw memory authoritative again:
aetnamem trial rollback

Activation does not discard native memory. Shadowing starts with a complete byte-for-byte baseline. Each observed native change is versioned, and before cutover AetnaMem takes a second complete snapshot of all files and empty directories in MEMORY.md and memory/. It refuses activation if the source changes while copying or any digest differs. It then deactivates the live copies, sets OpenClaw's native memory slot to none, disables the native session-memory writer, and points the bridge at the verified AetnaMem mirror. The bridge preserves OpenClaw's standard memory_search and memory_get tools. Results use stable aetnamem://record/... paths; exact reads of those records and digest-verified reads of frozen MEMORY.md / memory/*.md paths are audited through the engine's memory_get_record and memory_get_source MCP operations. After restart, activation inspects the loaded plugin runtime and automatically restores native memory unless both compatibility tools verify. AGENTS.md, TOOLS.md, identity, safety instructions and executable skills remain pinned in OpenClaw. Rollback restores and re-verifies the switch-time snapshot, returns active-period memories to OpenClaw, and preserves rather than overwrites any native paths unexpectedly recreated during takeover.

Activation also checks for configured native features the current takeover cannot preserve. It refuses the switch—with named reasons—when OpenClaw uses session-transcript search, extra memory paths, wiki memory, qmd, native multimodal indexing, memory-core dreaming, or the active-memory plugin. No configured corpus is silently dropped.

The 77-second real OpenClaw demonstration uses DeepSeek on an isolated agent. Its baseline answered “I do not know yet”; the AetnaMem arm answered “Kyoto” and used 12.9% fewer model tokens in that run. This is reproducible evidence for one demonstration task—not a promise that every workload improves by 12.9%.

The product promise is the safe evaluation path: see what AetnaMem mirrors, search it, inspect the evidence, and switch only when you are satisfied. Restore OpenClaw whenever you want.

Give your agent governed-memory skills

Install the AetnaMem engine, then install the repository's focused skills:

python3 -m pip install --upgrade aetnamem
npx skills add aetna000/aetnamem

The skill library gives Codex, Claude, and other compatible agents three plain-language workflows:

The skills do not replace the evidence engine. They call public AetnaMem CLI operations; the engine remains responsible for admission policy, SQLite storage, provenance, audit-chain verification, correction, and deletion receipts. OpenClaw or Hermes hooks are still required for automatic capture of authenticated turns.

Skills tell the agent how to use memory. AetnaMem proves what actually happened.

The repository includes Codex and Claude plugin manifests as well as the provider-neutral skills. See the Agent skills guide for trust boundaries, example conversations, wrapper commands, and validation.

What happens during a Safe Switch trial

For OpenClaw:

# Install the engine, then let it own the bridge installation.
python -m pip install --pre aetnamem==0.7.0a2
aetnamem --version
aetnamem openclaw install
aetnamem dashboard

--host auto means “detect exactly one supported agent executable on this machine.” It fails and asks you to choose if both OpenClaw and Hermes are installed; it never guesses. The beginner flow has no --subject: candidate memory is scoped internally as local-user. Applications with real tenant identity should continue using the explicit Python/MCP subject boundary.

The dashboard and CLI expose two states:

Internal evidence still distinguishes observations, retrieval decisions, and exposures, but those are not adoption modes. A missing, corrupt, or tampered state fails closed without injecting AetnaMem context.

aetnamem trial status
aetnamem trial activate  # asks you to type the detected host name
aetnamem trial rollback  # restore native memory and host configuration

Rollback restores the host configuration, native memory slot, session-memory hook, MEMORY.md and memory/. It does not delete trial evidence, reverse past agent outputs, or delete provider logs.

The dashboard is functional rather than experimental: it shows the active memory provider, the exact mirrored file manifest and SHA-256 digests, memory search results with source provenance, verification checks, and one activate/restore control. It does not claim that local answers improved. The checked-in OpenClaw benchmark is reference evidence, not a measurement of your installation. AetnaMem never spends provider money merely because it was installed.

Read the Safe Switch guide for host-specific details and exact rollback boundaries.

What is product, beta, and research?

This distinction is deliberate: “implemented” is not the same as “proven to improve every agent.”

Give OpenClaw four kinds of memory

You do not need to understand memory databases. First complete the Safe Switch trial. If you then want the advanced direct four-memory runtime, roll the trial back to its verified baseline before enabling the wizard configuration:

aetnamem trial rollback
aetnamem setup
openclaw aetnamem setup --single-user --subject you \
  --orchestrated --runtime-config ~/.aetnamem/runtime.json

The wizard walks through exactly ten short steps:

1. Explain what AetnaMem will add.
2. Choose a ready-made setup.
3. Name the person whose memories these are.
4. Name the agent.
5. Choose the private SQLite location.
6. Optionally point to OpenClaw SKILL.md files.
7. Choose where the configuration is saved.
8. Confirm the safety model.
9. Write and validate the configuration.
10. Print the exact OpenClaw and generic MCP commands.

For a scripted install with safe single-user defaults:

aetnamem setup --yes --preset starter --subject you \
  --agent openclaw-primary

Check the result before connecting an agent:

aetnamem runtime status --config ~/.aetnamem/runtime.json

Existing aetnamem mcp, Python Memory, and non-orchestrated OpenClaw setups continue to use the original semantic recall path. Four-memory orchestration is opt-in and falls back to legacy recall if the connected Python package does not offer the runtime tools.

The four memories in plain language

Imagine an agent completing a task:

What am I doing now?      → Working memory
What facts do I know?     → Semantic memory
What happened last time?  → Episodic memory
How should I do this?      → Procedural memory
                                  │
                                  ▼
                         One bounded context pack
                                  │
                                  ▼
                         OpenClaw or another agent

After the agent acts, AetnaMem can record the outcome reported by the caller. The generic CLI and MCP paths label this evidence caller_asserted; a trusted host integration must separately authenticate evidence before labeling it host_attested. Failed outcomes may create a quarantined lesson proposal for review. A skill can inform an action, but it never authorizes an action; optional Guarded Actions remain the approval boundary.

Ready-made configurations

The wizard creates normal JSON, but most users should start from a preset:

List or generate presets without the wizard:

aetnamem runtime presets
aetnamem runtime init --preset private --subject you \
  --agent openclaw-primary --output ~/.aetnamem/runtime.json

The same runtime works with Grok, Claude, DeepSeek, OpenAI, Ollama, or another model because coordination happens through generic MCP:

aetnamem runtime mcp --config ~/.aetnamem/runtime.json

See the four-memory runtime guide for the configuration, lifecycle, lesson review, and generic MCP tools. The four-memory ablation benchmark shows the complete runtime beside semantic-only and each leave-one-plane-out variant without making a model-quality claim.

What can you do with this repository?

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.

AI and agent use cases

Agent governance and operational use cases

Evidence-to-Decision and organizational use cases

Why this is more than another memory cache

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.

Try the local assistant today

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).

macOS

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.

Linux (Ubuntu/Debian and RHEL/Fedora/CentOS)

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).

Windows 10/11

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).

Using the app

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.

Data location and backup

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.
• Semantic search is optional and search-only. Its derived vector sidecar is plaintext SQLite, can be rebuilt, and never replaces canonical evidence. Every candidate is revalidated by subject, live status, and content digest. Local Ollama avoids content egress; a remote embedding endpoint receives indexed memory during build and each semantic query.
• 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.

Guarantee boundaries

Install & use

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 memories   ./memories.db --subject user-1 --all
aetnamem search     ./memories.db "preferred airport" --subject user-1
aetnamem trace      ./memories.db "preferred airport" --subject user-1 --output trace.json
aetnamem index build ./memories.db --subject user-1 --embedder ollama --model nomic-embed-text
aetnamem search     ./memories.db "departure location" --subject user-1 --mode hybrid
aetnamem index verify ./memories.db --subject user-1
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

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.

Collaborative decisions and EtD

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.

Use from agents (MCP)

aetnamem mcp currently serves 17 memory verbs 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_observe, 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_forget_artifact, 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.

memory_observe accepts a typed text observation from an image, audio, video, or document extractor and always quarantines it. memory_forget_artifact purges all indexed AetnaMem derivatives of one exact-byte SHA-256 while explicitly leaving the host's original file alone. See the multimodal observation guide. These two tools ship in Python v0.5.2; the 15 earlier MCP contracts are unchanged.

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 four-memory integration:

python -m pip install --pre aetnamem==0.7.0a2
aetnamem --version
aetnamem openclaw install

# Inspect the native mirror without changing prompts.
aetnamem openclaw memory status
aetnamem dashboard

# After inspecting and searching the verified mirror:
aetnamem trial activate

Activation freezes and relocates native supplemental memory, disables its duplicate retrieval/writer paths, and applies conservative AetnaMem context budgets. OpenClaw instructions, identity, safety, tools and executable skills stay pinned. 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.

Integrating with other agent frameworks

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.

Compliance posture

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.

Current semantic-memory storage pipeline

These L0–L3 labels describe the internal pipeline used by the semantic provider. They are not alternatives to the four runtime memory types above.

• 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().

How recall works

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

What the released core is and is not

Semantic 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 embedded four-memory runtime, setup presets, 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, remote memory-plane transport, 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.

Documentation

Adoption path

• Safe Switch guide — Beta: exact trial commands, host behavior, evidence meaning, activation, stop, and rollback boundaries.
• Agent skills guide — Beta: conversational governed-memory, audit, and Safe Switch workflows.
• OpenClaw setup and Hermes guide: host-specific installation and identity boundaries.
• Integration guide: complete CLI and MCP reference.
• 0.7.0a2 experimental notes: reliable daemon-lifetime access, direct system-browser launch, and fail-closed daemon startup when no protected URL is available; verified rollbacks can be activated again and interrupted switches surface a clear Restore OpenClaw recovery action.
• 0.7.0a1 experimental notes: staged CLI installation progress and visible dashboard progress during mirror refresh, activation, and restore.
• 0.6.1.1a6 experimental notes: dashboard search shows the best matching sentence or bullet instead of an unrelated full source chunk.
• 0.6.1.1a5 experimental notes: dashboard restarts now wait for and return the new one-time login URL instead of reusing a consumed code from the append-only daemon log.
• 0.6.1.1a4 experimental notes: two-state OpenClaw memory dashboard with exact source inspection, search, direct activation, and verified restore.
• 0.6.1.1a3 experimental notes: native-memory shadow, evidence review, verified takeover and rollback;
• 0.6.1 release notes: Safe Switch package versions, compatibility, validation, and known limits.

Released memory and audit surfaces

• Current capability status: the canonical released, beta, research, and planned boundary.
• Auditing guide and audit search/trace: investigate memory without knowing an internal ID and export text or JSON evidence.
• Multimodal observations: governed text observations with exact-artifact provenance; AetnaMem does not store or understand media bytes.
• Semantic investigator search — opt-in: local or configured embeddings for human investigation; default agent recall remains lexical.
• Data storage and backup and audit-log specification: persistence, recovery, replay, and verification boundaries.

Beta and advanced surfaces

• Four-memory runtime — opt-in: working, semantic, episodic, and procedural context orchestration.
• Graph memory — opt-in: derived, rebuildable relationship index with path evidence.
• Guarded Actions — Beta: approval and action evidence; not a universal MCP write interceptor.
• macOS desktop — Beta: local desktop workflow with platform-specific Keychain-backed sealing.

Research and experimental work

These components are checked in so their methods and evidence can be inspected. They are not part of the Safe Switch adoption claim:

• Memory Impact Lab — Research: balanced interventions, signed host verification, estimators, and a held-out policy gate. The 100-run Grok pilot is exploratory and incomplete; no general causal performance claim ships.
• Collaborative decisions / EtD — Research: a generic evidence-to-decision workflow, not a clinical or regulatory product claim.
• Governed outbound channels and inference engineering memory — Proposal: specifications, not implemented product surfaces.
• Architecture plan and roadmap: future work, not release documentation.

Benchmark

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

License

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.