Agent Memory
How an AI agent remembers the user it serves — what they said before, what they prefer, what context not to repeat — without that memory drifting the agent's behavior for everyone else.
What it is
Agent Memory is the part of an AI agent that lets it remember the user it's serving across sessions. Think of how a good colleague adapts: same person, different register for a new client versus a long-time peer; she remembers your project, your preferences, your past frustrations. An agent without memory restarts cold every conversation. An agent with memory can pick up where it left off — but the memory has to belong to the user, not poison the agent's behavior for everyone else who interacts with it.
Why it matters
Most production AI assistants today are amnesic on purpose: they forget everything between sessions to avoid the risks of getting things wrong. That's a safe default but it cripples usefulness — users repeat their context every conversation, the agent never builds rapport, and the simple act of saying 'use the format we agreed on last time' becomes impossible. Done right, memory is what makes an agent feel like it's actually working for you rather than starting over each time. Done wrong, it leaks one user's information into another's session, drifts the agent's personality, or remembers things the user wanted forgotten.
How it actually works
We model memory as four kinds with different lifecycles. Working memory holds the current task's context. Episodic memory keeps relevant past sessions. Semantic memory stores durable preferences and facts. Organizational memory is shared across a tenant when that's appropriate. Each entry carries a source (who or what asserted it), a confidence level, a decay policy, and an expiry. Retrieval at inference time scores entries by recency, relevance, and confidence; low-confidence stale entries are retired rather than poisoning future responses. Users can see what is remembered about them, correct it, and delete it — visibility and control are not features, they are requirements.
- Working memory: context for the current task
- Episodic memory: relevant past sessions
- Semantic memory: durable preferences and facts
- Organizational memory: tenant-shared context
- Per-entry source, confidence, decay, and expiry
- User-visible, user-correctable, user-deletable
What it works with
Sits on top of Synthetic Personality and below Closed-Loop Knowledge. Personality is invariant — what the agent is. Memory is per-user — how the agent relates to this specific person. Closed-loop knowledge is the system-wide layer — what the team learns from all interactions. The boundary matters: a single user should not be able to drift the agent's behavior for everyone, and what one user shares should not leak into another's session. Memory storage lives in Data Foundations with the same access controls as the rest of the substrate; retrieval runs through Vector Search.
Open questions we are studying
When does episodic memory help versus hurt — at what density of past sessions does retrieval start surfacing the wrong precedent? How should memory survive a model swap that changes the underlying embedding space? What is the right consent flow for organizational memory versus per-user memory? How do we audit a memory store for entries the user would correct if they saw them?
Prior art and adjacent work
Engages with MemGPT, LangMem, the long-horizon agent literature, and the broader research line on external memory for LLMs. Borrows from cognitive-science models of working / episodic / semantic memory. Adjacent to vector-search retrieval strategies and to the privacy work on user-controlled data stores.