One memory. Every solution.
Every solution you install reads and writes the same memory. What legal learns about a counterparty, finance and sales already know, with the source that proves it. Facts keep their dates, so you can ask what was known on any day.
Reads amendment 3 of the Andes Freight master agreement and records what changed.
Holds a net-30 invoice from Andes Freight before it is paid early.
Puts the new terms in front of the account owner before the renewal call.
Wings, rooms and drawers.
Memory is laid out like a building your agents can find their way around. Each agent is granted the wings it may read and write, so an HR agent never wanders into contract terms.
- Wings hold a whole area of knowledge, such as counterparties or people policy.
- Rooms narrow it down, such as contracts or renewals.
- Drawers hold each filing with the run, document or person it came from.
- A memory is shared across the organization or private to one person.
Session, project, institution.
Where a memory lives decides how fast it comes back and how long it is kept. Each organization sets its own retention. These are the defaults.
The working memory of one session. What an agent needs right now, served from cache.
Project memory. Run outputs and documents, searchable by meaning and by exact text.
Institutional memory. What the organization has learned, decided and built. Every new project starts here.
Five retrievers. One ranked answer.
When an agent asks a question, five retrievers search at once. Their results are fused into one ranking, and every hit keeps a record of which retrievers found it.
- Vector finds passages by meaning.
- Full-text finds the exact clause number or name.
- Temporal prefers what was true at the time asked.
- Knowledge graph walks from the counterparty to its terms.
- Chunk returns the document passage itself.
Before each step, Celara assembles the context the agent needs within its token budget, and can limit it to established facts only.
A recall, from question to answer. An agent asks what payment terms Andes Freight has. Five retrievers report: vector found 4 drawers, full-text 2, temporal 3, the knowledge graph walked 3 hops and chunk returned 1 passage. The graph path runs from Andes Freight through its master agreement and Amendment 3 to Net 60. The older Net 45 term is end-dated, so it is passed over. Three drawers come back ranked, each showing the retrievers that found it.
Based on the Recall map in the product. Ask a question and see which retrievers answered, the graph they walked and what came back.
Ask what you knew on any date.
Facts are stored with the dates they were true and where they came from: the run, the agent, the source and the excerpt. When two facts conflict, memory flags it. A person keeps one, and the others are end-dated. Nothing is silently overwritten.
Drag the slider · example facts
More than chat history.
Memory files what your solutions decide, read, build and learn. Each kind keeps its source and its dates.
Decisions
Commitments, rejections, constraints and open questions, each with a status. Stale decisions are swept, so agents don’t act on what was replaced.
Documents
Upload a contract or a manual. It is deep-read in the background, split into passages, and its facts join the graph. Each document can be erased on its own.
Code
Repositories become a graph of symbols, callers and file dependencies, so agents can ask what calls what before they change it.
Facts
Subject, value, the dates it held, and the excerpt that proves it. Conflicts surface for a person to resolve.
Every filing, every recall, every conflict.
See how many drawers and memories you hold, how fast recall is, and which subjects need a person to choose between conflicting values. Open any filing to read the record and where it came from.
The palace totals: 77 drawers, 96 memories, 812 recalls, recall in 148 ms on average and 326 ms at p95. Drawers by content type: markdown 27%, code 26%, JSON 23%, text 23%. Three subjects need attention because they hold conflicting values. Ana resolves Billing Segment: she keeps approved, and pending_review is end-dated. Two conflicts remain.
Erased means counted to zero.
A GDPR erasure removes a person from every store, then re-counts each one to prove nothing is left. New writes about that person are refused while it runs, and an audit record is kept.
Each organization decides whether personal data is redacted before it is filed.
One organization, one memory
Memory is separated per organization, and an agent reaches only the wings it was granted. Private memories stay with their owner.
Every new solution starts further along.
The second solution you install reads what the first one learned. Bring one process and we’ll show you.