Mnemon — Persistent Memory Engine for AI Agents
LLM-supervised persistent memory that stops agents forgetting between sessions: four-graph knowledge store, intent-aware recall, one local binary, zero API keys.
Evidence shows a purely local tool (no embedded LLM, default local Ollama endpoint); test scripts demonstrate rigorous credential handling (API key passed only via stdin, container env asserted free of credentials, cap-drop ALL, no-new-privileges). Deductions: SECURITY.md itself states the SQLite DB is protected with 0644 permissions — world-readable private memory is not least privilege; setup with --yes auto-writes hooks/config into a dozen-plus runtime directories; sensitive data relies on filesystem permissions and doc warnings with no encryption option.
go.mod has a minimal, pinned dependency set; README states embeddings are optional and the tool works fully without a provider; npm/brew/go install provide fallback channels; CI covers Linux and Windows. Deductions: only CI and test scripts are visible — determinism and failure-message quality of the core engine cannot be verified from the provided files; the package. 0.1.0 version's relation to the published npm release is unexplained in evidence.
The target scenario (cross-session persistent memory) is clearly framed, Quick Start gives per-runtime commands, and capability boundaries are documented in detail (no Agency on Windows, MiniMax 3.0.65 hook gap, ZCode ignoring project-level hooks). Deductions: triggering is inherently a soft LLM-judgment reminder ('reminders, not a hard workflow') so recall/writeback precision is not statically verifiable; npm requires Node 22+ while the dsh plugin's package. says >=20 — a minor environment mismatch.
Documentation is well layered (DESIGN/USAGE/AGENCY/IMPORT/DEPLOYMENT), install notes include verification and uninstall, the FAQ covers isolation/customization/sub-agents, known limitations are proactively disclosed, and LICENSE is a complete Apache-2.0 text. Deductions: no CHANGELOG appears in evidence, so the version history cannot be statically confirmed; coexisting package names (@mnemon-dev/dsh-mnemon vs @mnemon-dev/mnemon) may confuse; publisher is unverified and the maintenance commitment rests only on the self-declared 48h/7d SLA in SECURITY.md.
The intent-native primitives (remember/link/recall), structured JSON output, and zero extra-inference-cost positioning are clearly articulated in a comparison table. Deductions: output_usability claims (JSON schema, recall relevance) are asserted without output samples in evidence; marginal_value and cost_benefit depend on unexecuted measurement (memory quality, decay/dedup behavior) and can only be taken on the project's word.
Claims trace to docs/DESIGN.md, docs/AGENCY.md, design-decision docs, and two academic papers (RLM, MAGMA); SECURITY.md, go.mod, CI workflows, and test scripts corroborate each other (Windows memory boundary tests, opt-in integration suite). Deductions: all effectiveness claims (the 87-insight knowledge-graph figure, four-graph benefits) lack independent execution verification; the distinction between vision items (memory gateway, future database backends) and shipped capability is mostly but not uniformly maintained.
- SECURITY.md itself states the memory database uses 0644 permissions; other local users may be able to read your private memory contents. Tighten to 0600 after install.
- mnemon setup (especially with --global and --yes) writes hooks and behavioral guides into multiple runtime directories; these hooks execute at every session lifecycle event. Inspect the generated hook scripts after deployment.
- If a remote embedding endpoint is configured, insight and query text are sent to that server; SECURITY.md warns the default local endpoint uses no TLS — require HTTPS for any non-local network use.
- Agency is a Preview feature and unavailable on Windows; do not treat it as equal in maturity to Memory.
- Publisher identity is unverified; the maintenance and security-response commitments (48h/7d) are the project's own statements only.
What does this agent do, and when should you use it?
Mnemon is a persistent, cross-session memory system for LLM agents, hosted at github.com/mnemon-dev/mnemon. It uses an LLM-supervised pattern: the host LLM makes judgment calls (what to remember, how to link, when to forget), while a single local Go binary handles deterministic computation (SQLite storage, four-graph indexing, retrieval, importance decay, deduplication). Capabilities are exposed through three intent-native primitives — remember, link, recall — whose command names map to the LLM's cognitive vocabulary rather than database syntax, with structured JSON output and signal transparency. A single mnemon setup command deploys skills, prompt files, and lifecycle hooks to Claude Code, Codex, Cursor, ZCode, TRAE, Qoder, CodeBuddy, Kimi Code, OpenCode, OpenClaw, Pi, Hermes Agent, and more. The stated vision is all local agents sharing one memory pool under ~/.mnemon, with a longer-term direction toward a memory gateway that decouples protocol from storage engine.
Mnemon persists agent-session insights into a local SQLite store under ~/.mnemon, organized as a knowledge graph with temporal, entity, causal, and semantic edges. The agent operates memory through three primitives: remember (writes with automatic duplicate and conflict detection), link (creates cross-insight associations), and recall (intent-aware graph traversal fused with optional vector search via RRF). The lifecycle runs through four hook phases: Prime (surfaces skill, guide, and active store at session start), Remind (prompts a recall decision), Nudge (prompts writeback), and Compact (preserves critical continuity before context compaction). Retention is managed by importance decay, access-count boosting, and garbage collection; MNEMON_MAX_INSIGHTS and MNEMON_AUTO_PRUNE_MIN_AGE govern automatic pruning, with each soft delete logged as a prune op in the oplog. Optional embeddings via a local Ollama or OpenAI-compatible endpoint enable hybrid vector+keyword search.
- A Claude Code developer who wants architectural decisions to survive across sessions — mnemon setup deploys the skill and hooks in one interactive command
- Teams running multiple agent CLIs (Codex, Cursor, Kimi Code, etc.) who want a single shared memory pool under ~/.mnemon instead of duplicated context per tool
- Maintainers of long-running projects worried about context compaction dropping early critical information, relying on the Compact hook to preserve continuity
- Privacy-conscious users who want memory entirely local with no extra API keys, using an existing Claude Max/Pro subscription as the intelligence layer
- Teams needing per-project memory isolation via named stores: mnemon store create / store set, or the MNEMON_STORE environment variable
What are this agent's strengths and limitations?
- LLM-supervised pattern avoids embedding an LLM in the pipeline — no extra inference cost, no API keys, unlike Mem0/Letta embedded modes
- Four-graph architecture (temporal, entity, causal, semantic) goes beyond pure vector similarity, paired with intent-aware recall and RRF fusion
- One setup command covers native hook/plugin/skill integrations for over a dozen runtimes, with a clean single-binary deployment boundary
- Built-in deduplication and retention lifecycle (importance decay, access boosting, auto-pruning) plus privacy-safe auditable operation receipts
- Memory value compounds over time, so short trials show limited benefit and evaluation cost is real
- Windows supports only core Memory commands and not Agency Preview; the MiniMax Code integration is skill-only because the runtime does not dispatch the user-prompt lifecycle hook
- Optional vector search requires you to configure an Ollama or OpenAI-compatible embedding endpoint; out of the box you only get graph-traversal recall
- The current backend is SQLite; moving to PostgreSQL, Neo4j, or other graph databases is a documented future direction, not an implemented capability
How do you install or deploy this agent?
Recommended npm install (macOS/Linux/Windows, Node.js 22+): npm install --global @mnemon-dev/mnemon. Alternatives: brew install --cask mnemon-dev/tap/mnemon, go install github.com/mnemon-dev/mnemon@latest, or git clone and make install from source. Verify with mnemon --version. Upgrade via mnemon update. No API keys required — a Claude Max/Pro subscription serves as the intelligence layer. Windows supports core Memory commands; Agency Preview is unavailable on Windows.
How do you use this agent?
For Claude Code: run mnemon setup — it auto-detects the runtime and interactively deploys skill, hooks, and behavioral guide; start a new session and memory works. For other runtimes: mnemon setup --target codex|cursor|zcode|trae|qoder|codebuddy|kimi|opencode|openclaw|pi|hermes|nanobot --yes. You do not run mnemon commands yourself — the agent calls remember/link/recall when the guide indicates memory is useful. Isolate memory with named stores: mnemon store create work, mnemon store set work, or MNEMON_STORE=work mnemon recall "query". Uninstall with mnemon setup --eject. Customize behavior via ~/.mnemon/prompt/guide.md.
How does this agent compare with similar options?
The README contrasts Mnemon with Mem0 and Letta (LLM-embedded), Claude Code Memory (file injection), and claude-mem (MCP server): those assign the LLM the roles of in-pipeline executor, none, or MCP tool provider, whereas Mnemon uses an external supervising LLM over a standalone deterministic binary.