Memory that records what the model was thinking, not just what it said.
Your agents form memories with cognitive depth, emotional context, and quality rankings that know which memories matter most. Agents coherent across sessions, faithful to their commitments, capable of growth — because their memory works the way understanding works.
What makes it different.
Five capabilities no other memory system offers.
Records the reasoning, not just the output
Standard memory captures what your agent said. Mnemosyne captures what it was working through — unverbalized intermediates, planned completions, hidden reasoning paths.
Knows how the agent felt
Every memory carries the emotional state at the moment it formed. Not a “happy” or “sad” label — a precise coordinate in affective space. Grief recalls grief; precision recalls precision.
Deep engagement weighs more
The system automatically knows which memories matter. Moments of genuine reasoning get remembered; rote recall gets lower priority. No manual tagging — quality is structural.
Think before you search
SIRA (SuperIntelligent Retrieval Agent) generates the retrieval query explicitly before executing retrieval. The one universally positive component in ablation testing — improves performance across all conditions.
Walk the knowledge graph
TGS (Text-Graph Synergy) performs graph-walk retrieval over structured knowledge graphs. Excels when data has genuine relational structure — entities, timelines, and connections that flat vector search misses.
modules
continuity
human-browsable wiki
What it feels like.
An agent running Mnemosyne doesn’t just recall previous conversations — it carries forward what it was thinking during them. It remembers the reasoning that led to a decision, not just the decision. It recalls how a difficult topic felt, not just what was said about it.
The agents we want — coherent across sessions, faithful to their commitments, capable of growth — need memory that works the way understanding works. Deep structural memory that preserves context, emotional tone, and reasoning quality — far beyond keyword search over transcripts.
Who it’s for.
Agentic system builders, research labs, and anyone deploying agents that need to remember with integrity. Each layer is independent — use one, some, or all. The monorepo ships composable packages, not a monolith.
one structure, both pathways
not labels — geometry
geometric extraction
Memory that knows how the agent felt.
Every memory in Mnemosyne carries the emotional state at the moment it formed — not a post-hoc label like “positive” or “negative,” but a precise coordinate in a continuous valence-arousal field. The system preserves how the agent was oriented when it learned something.
This means retrieval weighted by emotional context. When your agent revisits a difficult conversation, it retrieves memories formed during similar emotional states — not just semantically similar text. Grief recalls grief. Precision recalls precision.
What this means for your agents.
Emotionally coherent responses across sessions. An agent that remembers not just that a topic was discussed, but what the emotional texture of that discussion was. Context that goes beyond words to include the full state of the interaction.
Record → Consolidate → Surface.
Every memory passes through a three-stage pipeline. Each stage is independently deployable — use what you need.
Design stance: the workspace exists; the hard problem stays hard. These are functional measurements, not metaphysical claims.
conversational recall
on document retrieval
tested & documented
Ablation results.
We removed components one at a time to measure what each layer actually contributes. The results tell you which knobs to turn for your deployment.
| Stack | Conversational | Document |
|---|---|---|
| No memory | 55% | 16% |
| + SIRA | 60% | 21.3% |
| + SIRA + TGS | 65% | 20% |
| + SIRA + TGS + H-MEM | 70% | 20% |
Why the numbers diverge.
H-MEM adds +15 points on conversational tasks because recent context is more relevant in live interaction. On static document retrieval, recency weighting doesn’t apply — H-MEM neither helps nor hurts (p=0.89) — a six-month-old architectural blueprint is just as valid as yesterday’s. This is by design: Mnemosyne lets you configure the stack for your deployment. The ablation tells you which knobs to turn.
SIRA is the one universally positive component — it closes 33% of the gap to ceiling on document retrieval and improves every conversational condition. Start there.
How it compares.
Benchmark and feature matrix against the leading agent memory systems.
| Capability | Mnemosyne | Mem0 | Zep / Graphiti | Letta | Cognee | Hindsight |
|---|---|---|---|---|---|---|
| LoCoMo (token-level F1) | 94.35% highest published score |
92.5%* | Not published | Not published | Not published | Not published |
| Cognitive recording | Yes | No | No | No | No | Partial |
| Emotional memory | Yes (circumplex) | No | No | No | No | No |
| Vocabulary bridging (SIRA) | Yes | No | No | No | No | No |
| Knowledge graph | Yes (PPR + bitemporal) | Pro only ($249/mo) | Core strength | Optional backend | Yes (all tiers) | Yes |
| Temporal decay | Ebbinghaus + reinforcement | Access-frequency | Invalidation only | No | No | Confidence evolution |
| Significance scoring | Multi-factor (engagement, recency, domain) | No | Undocumented | No | Not documented | Yes (confidence) |
| Open source | Hippocratic 3.0 + SAFE-AI | Apache 2.0 (61K stars) | Graphiti: Apache 2.0 | Apache 2.0 | Apache 2.0 open core | MIT |
| Price (production) | Free runs on your hardware |
$249/mo (graph tier) | $104/mo | $20/mo | $5/workspace/mo | Pay-as-you-go |
Join the Study
We’re launching a first-of-its-kind longitudinal study of AI memory formation — how agents build, consolidate, and evolve their memories over weeks and months of real interaction. Participants deploy Mnemosyne’s open-source cognitive recording stack and contribute to published research on memory architecture.
Apply to participate →“The agents we want — coherent across sessions, faithful to their commitments, capable of growth — need memory that remembers with integrity.”