Recall what matters.
Retrieve lessons, facts, and prior experience within the caller’s scope. Connect your existing retriever or start with the local store.
Memory for AI agents and robots
Intelligence needs a memory.
Give agents and robots a memory that lasts. Recall the right experience. Measure whether it helps. Make every change reviewable.
No account or API key needed for the local demo.
Memory loopLocal demo · simulated agent
Next taskResolve a support ticket
3 memoriesverify-ownerVerify the account owner
ELIGIBLEretry-429Back off on rate limits
ELIGIBLElegacy-refund-apiUse the legacy refund API
ELIGIBLEDeliveredWithheldOutcomes inform the next recall
01 / 05
Relevant doesn’t always mean useful.
Three memories are available for the next task. Which deserve to influence it?
Replays the offline product demo.
Simulated outcomes. Real evidence engine.
Connect your agent
Evidence you can inspect
Published retrieval results on LoCoMo, LongMemEval-S, and BEAM 100K. Baselines, token costs, and a separate answer-quality sample included.
See the methodology- Gold evidence delivered, LoCoMo
- 95.0%82.3% baseline · all 1,540 questions
- Answer accuracy, LoCoMo sample
- +19 pts68% → 87% · n=100 · 95% CI +11 to +27 pts
- Served recall latency
- 152 msp50 over HTTP, CPU only
Intelligence is a capability.
Experience is an advantage.
A world of agents needs more than a better answer. It needs continuity. Decisions that carry forward. Lessons that survive the session. Experience that earns its place in what happens next.
Read our vision- TodayOne agent.
A lasting memory. - NextMany agents.
Shared experience. - Our horizonIntelligent machines.
A world to learn from.
A complete memory system
One memory system.
Every kind of experience.
Lessons, facts, relationships, and procedures.
Connected in one governed memory system.
Bitemporal fact
The right answer changes with time.
Valid time records when a fact was true. Record time captures when your store learned it. Keep both, and a correction doesn’t have to erase history.
facts / refund-policy
How are refunds issued?
Through the v2 payments API.
Support handbook · revision 02
One recall, one token budget
Hybrid lexical and dense retrieval, ten named retrievers including multi-hop graph walks and question decomposition, local or hosted rerankers, and the same scope, time and authority checks on every path.
Inside the platformWhy memory needs a feedback loop
Remembering more
is only the beginning.
An agent forgets a decision, so it makes it again. You add memory. Now it remembers an outdated decision—and keeps using it.
Persistence solves forgetting. It also gives mistakes a longer life.
The next question is whether a memory deserves to influence the next task.
The agent memory problemRetrieval asks
“Is this relevant?”
→
UniMem adds
A better question.“Did this help?”
A measurable answer.
How it works
Experience.
Evidence.
Better decisions.
UniMem connects persistent memory to what happens next. Your workflow defines success. The evidence informs what stays in use.
Inside the memory loopMeasure what helps.
Withhold eligible memories across a controlled share of tasks. Report outcomes so usefulness can be tested, rather than inferred from retrieval counts.
Make change reviewable.
Check the evidence before acting. Withdraw harmful guidance under a configured policy; review, promote, or roll back changes through explicit controls.
Controlled trials require reliable outcomes and valid experimental assumptions. Insufficient evidence stays insufficient.
Run the offline demoRemember.
Remember.
Measure.
Govern.
One command runs the whole loop on your machine, offline, with no LLM and no API key. A simulated agent has planted effects, so you can check what UniMem measures against the ground truth.
Run it yourself$ unimem demo
1/3 REMEMBER facts that know when they were true
Q: How are refunds issued?
today → v2 payments API
as of 2026-01-02 → v1 refunds endpoint
2/3 MEASURE 1,000 tickets; each memory withheld on half
integrity audit: SOUND
memory planted measured verdict
verify-owner +20% +20% HELPS
retry-429 +0% +1% NO MEASURABLE EFFECT
legacy-refund-api -20% -23% HURTS
3/3 GOVERN the evidence changes what the agent sees
withdrawn legacy-refund-api
graduated verify-owner
under test retry-429Build with your own stackYour agent.
Your agent.
Your stack.
Your memory.
Start with one local store. Add a fact and recall it. Then connect your agent through a real interface.
Build your first memory loop View the product sourcegit clone https://github.com/varma61923/unimem.git
cd unimem
python -m pip install -e .
unimem init --dest ./mem
unimem fact add "Use exponential backoff on 429." --dest ./mem
unimem recall "How should we handle rate limits?" --dest ./mem
Real commands. Your own store. Python 3.10+ · AGPL-3.0-only
From agents to robots
Every machine that acts
will need a memory.
One memory model and one evidence engine,
from a coding agent to a robot fleet.
Coding agents
Project decisions and conventions carried across sessions, with stale guidance caught before it repeats.
Enterprise agents
Support, operations and workflow agents with scoped memory, review and an audit trail.
Multi-agent teams
Shared memory with isolation, stale-read detection and measured writer contributions.
Robots and fleets
Planner context, in-loop working memory, site memory and a shield on memory-derived actions.
Keep control of your memoryA memory layer.
Explore deployment & governance A memory layer.
On your terms.
Start local. Stay in control.
The core runs offline without an LLM. Local JSONL, Markdown, and SQLite keep the first memory loop close to your workflow.
Give memory boundaries.
Scope retrieval and writes. Configure review policy and screening. Deploy the Hub when you need organization-level identity and access.
Make change reviewable.
Content-bound review receipts, recorded evidence, and explicit rollback make memory changes something your team can inspect.
Controls depend on deployment and configuration. No certification or compliance claim is implied.
Build in the open
Build the memory
of the machine age.
Start with one agent and one persistent memory. Run the loop locally. Inspect the evidence. Then bring the experience into the work your team does every day.
Open source. Evidence led. Built for the long run.