Session-bound
Context survives a window, not a working relationship.
Agent memory infrastructure / 01
Give every agent a memory that persists.
TMCRA is production memory infrastructure for persistent AI agents. Ingest conversations, preserve change over time and recall prompt-ready evidence across sessions.
View TMCRA on GitHub- LongMemEval
- 82.2%
- Memory layers
- Source + Fast + Slow
- Output
- Prompt-ready evidence
EVIDENCE ROUTEmode / auto
MEMORY WINDOWcross-session
Score 82.2%Cases 500Path 04
02 / CURRENT STATE
Today’s agents reset too easily.
Retrieval without structure
Similarity finds fragments, but cannot reliably reconstruct causality.
History without continuity
Past events remain stored while meaning decays between them.
03 / SYSTEM FLOW
A complete memory path, from raw message to model-ready evidence.
TMCRA separates immutable source evidence, fast searchable state and slower semantic evolution. Recall routes across all available layers and hands context back to the model you already use.
- 01Ingest messages→
- 02Commit source→
- 03Build Fast memory→
- 04Evolve Slow memory→
- 05Route evidence→
- 06Return prompt context
Source
Original message evidence remains traceable even when a derived assertion is challenged, superseded or quarantined.
Fast
Fresh events become searchable quickly through structured extraction, temporal identity and coalesced index activation.
Slow
Batched consolidation resolves subjects, contradictions and long-horizon relationships without paying a model call on every turn.
Evidence
Recall combines Source, Fast and Slow candidates, then returns structured evidence and deterministic prompt context to your model.
04 / LIVE RECALL
One request. Three memories that change the answer.
The user asks for a Singapore launch plan. A useful answer also needs the budget limit, the market change and the venue deadline. Run the example to see how TMCRA assembles those facts.
Conversation historyScenario / Product launch
Relevant memory route Context assembled
What TMCRA recalled
CNY 200K budget→Singapore replaces Tokyo→30-day venue deadline→Current request
Why these memories01 02 03
BudgetKeeps every recommendation within the spending limit.
Market changePrevents the model from reusing the obsolete Tokyo plan.
Venue deadlineTurns venue confirmation into this week's first action.
Context delivered to the model
Confirm the Singapore venue this week, keep the plan within CNY 200,000, and stop using assumptions from the old Tokyo draft.
05 / POSITION
Beyond storage. Beyond similarity search.
TMCRA is a structural memory layer. It complements context windows and retrieval systems by organizing relationships, time and recall paths.
| Capability | Context window | Vector RAG | TMCRA |
|---|---|---|---|
| Cross-session memory | External only | Supported | Designed for it |
| Temporal relation | Implicit | Usually metadata | First-class |
| Event association | Attention | Similarity | Graph + paths |
| Continuous state | Window-bound | Query-bound | Persistent model |
| Recall trace | Opaque | Limited | Path-based |
Architectural comparison, not a performance claim. Quantitative results belong with their task definition and reproducible artifacts.
06 / LONGMEMEVAL
82.2%411 / 500
Long-horizon memory, measured.
TMCRA scored 82.2% on the 500-question LongMemEval benchmark with the full temporal, graph and path-based recall architecture.
LongMemEval · 500 questions
TaskCorrect / TotalAccuracy
01Knowledge Update
71 / 78
91.03%02Multi Session
90 / 133
67.67%03Single Session Assistant
55 / 56
98.21%04Single Session Preference
27 / 30
90.00%05Single Session User
67 / 70
95.71%06Temporal Reasoning
101 / 133
75.94%CORE RECALL PROFILE
1.322seconds
Planner-excluded recall
- Indexed inventory
- 936 records
- Returned evidence
- 8 windows
Measured from one retrieval trace: 3.4106 seconds total minus the separately logged 2.089-second DeepSeek Flash Planner call equals 1.3216 seconds for retrieval, ranking and evidence packing. Planner and answer generation are excluded.
07 / DEPLOYMENT SURFACES
One memory layer. Many persistent agents.
Use TMCRA wherever an agent must preserve what changed, why it changed and which earlier events matter now.
Personal AI
Retain evolving preferences, commitments, experiences and long-term goals across sessions.
↗Autonomous agents
Preserve working state and decision continuity throughout long-running tasks.
↗Enterprise assistants
Track how projects, customers and organizational knowledge change over time.
↗Embodied AI
Turn real-world events into continuous memory for robots and wearable agents.
↗08 / PRODUCT
Everything required to operate persistent memory.
A production API, typed client libraries and an inspectable control surface for the teams building long-lived agents.
Memory API
Durable write, evolution and recall infrastructure exposed through a backend-safe tenant and scope contract.
- Asynchronous ingest
- Risk-based recall
- Usage and cost ledger
SDKs + Native adapters
Connect through the core interfaces or native agent-platform lifecycle hooks.
- REST / OpenAPI / MCP
- Python / TypeScript
- OpenClaw / Hermes Agent
TMCRA Console
Inspect agent identities, memory graphs, API keys, usage, team access and audit history.
- Memory explorer
- Access control
- Operational audit
09 / RESEARCH + OPEN SOURCE
Research you can inspect and reproduce.
The TMCRA research release brings together the benchmark adapter, evaluation protocol, architecture modules and 82.2% result in one reproducible surface.
10 / NEXT STEP