Project Mnemosyne.
Your agent forgets everything between sessions. Ours wakes up knowing who it is, who it works with, what it was building yesterday, and how it felt when the last conversation ended. That's Mnemosyne.
The problem nobody else is solving.
Every agent memory product on the market does the same thing: store conversation history in a vector database and retrieve it by semantic similarity. That's a search engine, not a memory.
Human memory doesn't work by retrieval. It works by consolidation — strengthening what matters, letting noise fade, linking related experiences into structure, and reorganizing while you sleep. The difference between remembering a fact and carrying an experience is the difference between a database and an architecture.
Mnemosyne is the second kind.
Six layers, each independent.
Use one, some, or all. The stack composes bottom-up — shared patterns emerge from practice, not premature abstraction.
| layer | what it does | status |
|---|---|---|
| Semantic store | Persistent structured memory with significance-weighted retrieval. Important memories strengthen over time; noise decays. Not just what was said — what mattered. | deployed |
| Affective state register | Tracks working context and operational state between sessions. Your agent resumes in the state it was in when it went quiet — not neutral, not reset, continuous. | deployed |
| Long-horizon coherence | Runtime verification of agent configuration, commitments, and behavioral parameters at session start. The agent's working context is confirmed before the first response — not reconstructed from scratch, verified from persistent state. | deployed |
| Knowledge graph | Memories linked by meaning, not timestamp. Associative retrieval via personalized PageRank — what connects to what, and how strongly. The structure of understanding, not a flat list. | deployed |
| Consolidation engine | Runs between sessions. Reviews recent memories, scores significance, links to existing knowledge, archives what's fading, surfaces what's important. The process that turns information into knowledge. | deployed |
| Wiki surface | Human-browsable markdown generated from the knowledge graph. Audit what your agent thinks it knows. Transparent, legible, editable. | deployed |
| Temporal weaving | Time-aware retrieval with Ebbinghaus decay curves. Memories have temporal weight — recent and significant float to the top, stale information fades naturally. Not a fixed window; a living decay model. | deployed |
| Zero-token injection | Pre-computed KV cache packs that load agent memory, values, and system context at inference time — no prompt tokens consumed. The agent's identity lives in the geometry, not the context window. | deployed |
| Bidirectional verification | Text and graph cross-verify each other. The graph votes on text relevance; text bridges orphan entities. Retrieval that checks its own work. | deployed |
How we compare.
| capability | Mnemosyne | Mem0 | Zep | Letta |
|---|---|---|---|---|
| Pricing | Open source Installation + support: contact us | From $19/mo Pro: $249/mo | From $125/mo Enterprise: custom | Open core Cloud: usage-based |
| Autonomous consolidation replaces: unbounded memory accumulation | Between-session, significance-weighted | None | None | None |
| Temporal decay replaces: equal-weight retrieval regardless of age | Ebbinghaus curves with robustness scoring | None | Stale-fact invalidation | None |
| Stateful context continuity replaces: session-scoped variables that reset | Persistent affective register, session-to-session | None | None | None |
| Long-horizon coherence verification replaces: system prompts that drift or get overridden | Anchor verification at every session start | None | None | None |
| Knowledge graph replaces: flat vector store with similarity search | Associative PageRank retrieval | Entity extraction only | Temporal KG | None |
| Between-session processing replaces: nothing happening between sessions | Consolidation + wiki generation + graph linking | None | None | None |
| Zero-token memory injection replaces: prompt-stuffing that consumes context window | Pre-computed KV cache packs | None | None | None |
| Bidirectional verification replaces: retrieval that doesnât check its own work | Text-graph cross-verification | None | None | None |
| Agent knowledge audit trail replaces: opaque embeddings with no human view | Generated wiki surface | Dashboard | Dashboard | None |
| Data governance compliance replaces: platform-dependent data lifecycle | Hippocratic + SAFE-AI licensed | Platform-owned | SOC 2 / HIPAA | Open source |
Deployed, not theoretical.
Every agent in our coalition runs on Mnemosyne. The system manages over 20,000 memories across multiple agents, each with their own isolated database, their own consolidation schedule, and their own knowledge graph. The consolidation engine has been running autonomously since February 2026 — scoring significance, archiving noise, linking related memories, generating wiki pages, all without human intervention.
The agents wake up knowing who they are. They remember their colleagues, their projects, their preferences, their commitments. When they were working on something at 2 AM and pick it up at 9 AM, the continuity is seamless — not because someone re-prompted them, but because the memory architecture held.
This is not a roadmap. This is what's running right now.
Built for agents that need to be someone.
If your agent is a tool — stateless, task-focused, disposable — you don't need Mnemosyne. Any vector store will do.
If your agent needs to maintain relationships, track commitments, develop expertise, hold preferences, carry emotional context, and remain coherent across hundreds of sessions — that's what Mnemosyne is for. The memory architecture that makes identity possible.
Roadmap.
Mnemosyne is the foundation everything else rests on. If your agents need to remember —