Remembers people
Keeps durable facts about each person it works with, what they prefer, what they decided, who they are, scoped per agent so trust boundaries stay clear.
‹ qOS
Agent Memory
Agents that learn from doing the work.
Most assistants start cold every session: they read the current thread, guess, and forget. Agent Memory is the layer that changes that. It is a layered memory system that records what people told it, what it learned about the companies and deals you work on, and how it solved multi-step tasks, then consolidates all of it while the agent is idle, so every conversation compounds on the last.
7 layers working memory up to governance
Sleep-time consolidates between sessions
You control it see, correct, delete, freeze
Why Agent Memory
Every session started from zero.
A chatbot with tools reads the current thread, makes its best guess, and forgets the moment the conversation ends. Bolt a vector store on the side and you get search over old transcripts, not learning: the agent re-asks what you already told it, loses the thread of a decision made last month, and never gets better at the work.
Agent Memory is built the other way around. It distills durable facts from each conversation, reconciles them against what it already knew, and consolidates the result while the agent is idle, so context compounds instead of resetting.
Capabilities
What it can do.
Builds institutional knowledge
Learns about the companies, accounts, deals, and projects you work on, and shares that knowledge across every agent on the team, so asking what you know about an account is one question, not an archive dig.
Remembers how it did the work
Records the steps and tools behind a successful multi-step task, so the next report follows the path that worked and skips the dead ends.
Consolidates while idle
A sleep-time process distills new facts and weighs each one against what it already knew, judging whether it refines, contradicts, or duplicates an existing memory before writing, then merges related facts into one clean record. The difference between learning and just storing transcripts.
Recall that blends keyword and meaning
Hybrid search over everything it knows, keyword plus semantic, refined by a re-ranking model, so the right memory surfaces even when you phrase it differently.
Surfaces context on its own
On every turn it quietly pulls the few most relevant facts into view, so the agent never misses what the team already settled, without being told to look.
Never overwrites, always traceable
New facts supersede old ones with a backward link instead of deleting them, so every memory carries its source and you can always see when and why it changed.
Workspace-aware from day one
Keeps a live briefing of the workspace itself, active people, key channels, recurring themes, recent decisions, so an agent arrives knowing the room.
Yours to see, and to forget
Anyone can ask what an agent knows about them, correct a fact that changed, or have it forgotten for good, a real deletion on request, not a hidden flag. Built for the EU AI Act and right-to-be-forgotten from the start.
Built different
Why it is not a vector store on the side.
-
Layered, not bolted on
Seven layers (L0 to L6) from the live thread up to governance, each with its own read and write path, feeding each other through sleep-time workers. Memory is part of the architecture, not a database stapled to the edge.
-
Per-agent boundaries, by design
What one agent knows about a person is a separate record from what another knows, so the question of which agent may see which memory has a real answer. Cross-tenant memory does not exist, a hard rule, not a setting.
-
Entity memory compounds across the team
Institutional knowledge about an account or deal builds across every agent in the workspace, the single biggest lever for a team that lives in B2B.
-
Procedural memory is the moat
Knowledge of how to do the work gets stronger with use and is not transferable, so a fresh competitor has years of catch-up, not a feature to copy.
-
Governance is first-class
Every memory action, what the agent learned, reconciled, surfaced, or forgot, lands in a searchable, provenance-stamped audit trail, identity-scoped on every read and write. The traceability the EU AI Act now requires, built in rather than retrofitted.
-
Built on systems we own
A durable store of record holds the source of truth; Agent Search indexes it, Agent Cache keeps the hot set in microseconds, Agent Storage archives it. The whole memory layer rides the rest of qOS.
By the numbers
How recall stays sharp.
Hybridkeyword plus semantic recall
Re-rankeda cross-encoder picks the best
Every turnrelevant facts auto-surfaced
Append-onlynever deleted, always audited
Agent-first
Memory the agent composes against.
Memory is not a private feature hidden inside the runtime; it is exposed as standard tools the agent uses the same way it uses everything else. It can search what it knows, ask what it knows about a person or a company, record a fact it just learned, or correct one that changed, all through the same protocol, all identity-scoped and audited. And because the most valuable memories are the ones the agent would never think to look for, the platform surfaces them automatically on every turn.
- Pull on demand: search memory, look up a person or an entity, review how a task was done before.
- Write deliberately: assert a fact, supersede an outdated one, all preserving the audit chain.
- Hand off safely: a peer agent gets a signed, time-boxed grant to a slice of memory, not the whole store.
Put Agent Memory to work.
See HQ running in your own Slack or Teams, on the operating system we built for agents.