How to build AI agent memory

Replaying conversation history is the default memory model, and it fails in three ways at once. Cost: context grows with every turn, and you pay for all of it on each call. Recall: the fact you need is buried in turn 14 of a session from last week, and attention over a very long context is unreliable. Structure: a transcript records what was said, not what is true — nothing marks which statements were corrected, superseded or acted on.

Summarisation trades those problems for a lossier one: the summary keeps what seemed important at summarisation time, not what the next question will need.

What an agent actually needs to remember is relational. The user's preference relates to a project; the incident relates to a service, which depends on two others; today's decision supersedes last month's. Store those as typed relationships and recall becomes a query: "everything known about the entities in play, newest first, with sources." No similarity guessing, no replay — a traversal that returns the same answer every time it's asked.

Four node types cover most agent-memory needs. Entity: the durable things — people, services, accounts, projects. Fact: a statement with an observed_at timestamp, linked ABOUT the entities it concerns. Source: where the fact came from — a conversation, document or API call — linked by SOURCED_FROM. Session: the run that produced the observations, so memory can be scoped or expired per session.

Corrections don't delete: a new fact with a SUPERSEDES edge to the old one preserves the history while queries default to the newest. That gives you time-travel and auditability for free.

After each meaningful step, the agent (or a cheap extraction pass) turns what happened into entities and facts and writes them with MERGE — get-or-create semantics, so observing the same entity twice converges instead of duplicating. Writes in CognoDB are ACID, so a fact and its edges land atomically.

At the start of a turn, resolve the entities the query mentions (exact match, or CognoDB's built-in BM25 full-text search for fuzzy names), then walk outward: facts about those entities, their sources, and the entities one hop further that recent facts connect to. The result is a compact, citable context block whose size is set by the traversal you wrote — not by how long the agent has been alive.

Memory topology is an architectural choice, and instances cheap enough to provision per agent make both patterns practical. Isolated: one CognoDB instance per agent, session or tenant — separate credentials, separate failure domains, nothing to leak between customers. Shared: a fleet of agents reading and writing one graph as a blackboard, where one agent's observation is immediately traversable by the others.

Either way the agent connects with any Bolt driver, or queries its memory directly through the built-in MCP server — the agent reads the schema and writes its own Cypher, and the query it ran is the citation for what it recalled.