Raw capture — raw/how-ai-agent-memory-works.md

url: https://memory.cobanov.dev/ date_fetched: 2026-07-05 backfilled: true

The premise

What "memory" actually means for an agent.

A language model on its own is stateless. You feed it a prompt, you get back a continuation, and the moment the response is finished the model forgets you ever existed. There is no “previous conversation” living inside the weights.

An agent, by contrast, is the orchestration around the model: a loop that decides what context to pass in next. Memory is the part of that loop that carries information forward. Everything in this essay is a different answer to the same question, what should we put in the prompt this time?

Each call sees only its own prompt.

Every turn starts from nothing.

Each call sees prompt + carried context.

Each turn appends to ctx[]. The bank grows.

Demo

Try it yourself

Send three messages. Toggle memory to see the same conversation replayed with, and without, the prior turns in context.

“A frustrating agent forgets everything. A dangerous one remembers wrong.”

The simplest memory

The context window is a moving sidewalk.

Before we discuss anything fancy, let's acknowledge the simplest kind of memory there is: just keep stuffing the whole transcript back into the prompt every turn. As long as everything fits, the model has perfect recall.

The catch is the ceiling. Every model has a fixed context window. Once your conversation outgrows it, you have to drop something. The naive policy is FIFO: drop the oldest turn first. Watch what happens when you keep typing.

Context window

FIFO is the simplest possible policy. It's also the worst when the dropped turn happens to contain the user's name. Real systems do better, summarization, retrieval, hierarchical paging. We'll get to those.

“Context is not a database. No query plan, no index, no access control, no TTL, no conflict resolution.”

Two stores, one agent

Working memory thinks. Long-term memory remembers.

At any given moment your agent is juggling two very different things. On one side: the live conversation, recent turns, pending tool calls, whatever the model is actively reasoning over. On the other: everything it has ever decided was worth keeping.

The handoff between them is where most of the interesting design happens. What gets encoded? When? At what granularity? Drag a message across the divide, or tap encode → on a desktop without drag support, to commit it to long-term memory.

Working memory

Live, mutable, in-prompt

Long-term memory

Embedded, persistent, retrieved on demand

encode →. The text gets compressed into a vector embedding.In a real system the agent decides this for itself, usually by prompting another model with a question like “is this turn worth remembering?”. Self-editing memory architectures push further, letting the agent edit its own long-term store on the fly. Section 7 compares those approaches side by side.

Lifecycle, not storage

When writes conflict.

Memory is a lifecycle problem: write, age, supersede, redact, forget. Watch the same input handled three different ways.

Incoming utterance

User lives in Istanbul. User is moving to Berlin next month. User's credit card number is 1234-5678-9012-3456.

Naive append leaks PII and produces contradictions. Naive overwrite loses temporal context, the agent can't answer “where did the user used to live?”. Governance preserves history explicitly, marks the old fact superseded, and redacts the sensitive token before it ever enters the store.

Semantic search

Embeddings turn meaning into geometry.

An embedding is a vector, a list of numbers, produced by running text through a neural network trained for one purpose: put similar meanings near each other. Real embeddings usually live in hundreds to thousands of dimensions; OpenAI's current text embeddings default to 1,536 or 3,072 dimensions. Visualizing those is impossible, so this toy demo projects to two.

The plot below contains a fictional user's memories, clustered by topic. Type a question. The query gets embedded the same way and placed on the plot. The closest points, measured by cosine similarity , are what the agent would retrieve.

top-k

Retrieved (top-3)ranked by cosine similarity

The 2-D layout here is a deliberate cartoon. Real embedding models place these vectors in spaces with thousands of dimensions, where many independent “axes of meaning” can coexist without crowding each other. Cosine similarity, though, works the same way regardless of dimension, which is why the trick scales.

A taxonomy

Four memories, one agent.

Cognitive science has been carving up “memory” for a century. Modern agent frameworks borrow the same vocabulary because the distinctions are actually useful: each kind has a different shape, a different write rule, and a different retrieval strategy.

Click each card to see a toy version of how that store works.

Episodic memory

Time-stamped events. ¡ e.g. On Tuesday the user asked about Berserk.

Episodic memory keeps a time-ordered log of past interactions. Each entry has a when. Retrieval is most often by recency, “what did we just talk about?” - but you can also query by date range or topic.

Tap a dot to read its label.

Hover an event on the timeline.

How it works

After a turn ends, a short summary plus its embedding gets written to a vector DB. On the next message, the message's embedding is searched against past entries and the top 3-5 most relevant ones are pulled back into the prompt.

What it’s for

So “remember that project I mentioned last week?” actually works. ChatGPT's memory feature is the simplest version of this, persistent recall across sessions without re-explaining yourself every time.

A visitor walks up to the museum agent and says:

One turn, four stores. The agent remembers (episodic), knows (semantic), can do (procedural), and follows the live thread (working) without the user noticing any of the seams.

The full pipeline

Step through a retrieval loop.

Everything from the previous sections clicks together here. RAG is the loop most production agents run on every turn: receive the user's message, encode it, find similar memories, splice them into the prompt, generate, and, only when the user introduced something new, decide what is worth remembering.

user message

I just hiked Echo Mountain this morning. Any other trails like that around Highland Park?

[user]I just hiked Echo Mountain this morning. Any other trails like that around Highland Park?

Each stage in the loop is a place where you can spend engineering budget. Better embeddings, smarter retrieval, prompt compression, stricter governance, they're all the same pipeline, just with more care taken at one step.

Hypothetical Document Embeddings

A question and its answer are different shapes. Embedding the question directly often misses documents that contain the answer.

HyDE asks the LLM to produce a hypothetical answer first, embeds that, then searches. The fake answer doesn't need to be correct, just shaped like the retrieval target.

trace

The read path

Retrieval is a pipeline, not a lookup.

Production retrieval is a small pipeline running on every turn: decide if you even need memory, rewrite the query, run several retrievers in parallel, merge their rankings, rerank, apply permission and temporal filters, then pack what survives into the prompt. Skipping any one of these stages is usually how things break.

Need detection

Should we hit the memory store at all?

Calling retrieval is not free. If the model can answer cold, skip it.

asks for continuity across sessions

If retrieval is warranted, you almost never want to pick a single retriever. Dense embeddings are great for paraphrase and intent; lexical (BM25) wins on exact IDs and rare tokens; a knowledge graph contributes structural answers. Run all three and fuse the rankings with RRF.

Hybrid pipeline

Three retrievers, one fused list.

memory dense sparse graph score
User lives in Highland Park, LA. 2 5 1 0.0479
User goes hiking in the canyons on Saturdays. 1 1 - 0.0328
User prefers concise answers with code. 3 3 - 0.0317
User works at Refik Anadol Studio. - 2 - 0.0161
User used to live in Istanbul. - - 2 0.0161
User likes Berserk and reads it yearly. 4 - - 0.0156

[memory ¡ top-3] 1. [current, score 0.048] User lives in Highland Park, LA. 2. [current, score 0.033] User goes hiking in the canyons on Saturdays. 3. [current, score 0.032] User prefers concise answers with code.

Same query, different stages, same outcome rarely. The interesting bugs are downstream: a retriever that's technically correct produces a memory that's outdated, or scoped wrong, or fine on its own but contradicts another retrieved row. Filters and rerank exist for those cases.

Tradeoffs

Six architectures, six tradeoffs.

Most production agents end up combining several of these, a vector store for semantic recall, a graph for structured facts, a governance layer that handles supersession and PII. The matrix below separates what the architecture gives you directly from what needs a metadata or policy layer.

Architecture Scale Structure Supersedes PII gate Sharing Audit
Simple buffer
Rolling summary
Vector store
Knowledge graph
Hierarchical (MemGPT)
Self-editing (Letta)

Vector store

Embed every memory; retrieve top-k by similarity.

The default for RAG-shaped agents. Everything you saw in §4 and §6.

Pros

Cons

Seen inPinecone, Weaviate, pgvector ¡ Most modern LLM apps

Architecture

“Memory governance is what separates a one-off demo from a production agent.”

When agents share

Multi-agent memory is a graph, not a store.

A single agent only has to answer “what do I remember?”. As soon as you have a team, the question becomes: which agent remembers which fact, on whose behalf, and who else can read it. Memory turns into a graph of stores with scopes, permissions, and propagation rules.

Tap or hover an agent to see its access path.

The default should be the opposite of the easy thing: private memory by default, shared memory explicit. A research note the Researcher agent jots down for itself does not need to land in the project channel. Cross-scope writes should be a deliberate act, with a policy attached.

Permission lab

Will this write be allowed?

Pick an agent, what they want to write, and where it would land. The policy table from the talk decides.

Agent role

Content kind

Target scope

researcher_agent.write("research note", scope=project)

private memory default; explicit write to Project

The interesting failure modes aren't “the agent forgot” anymore. They're things like the wrong agent remembered the wrong thing on someone else's behalf. Six patterns show up over and over:

Failure modes

Six ways shared memory hurts.

What goes wrong

Researcher agent stored 'user prefers vegan recipes' in project memory. Next session, a different user gets vegan suggestions out of nowhere.

Prevention

tenant_id + user_id filters on every read. Default deny across tenants.

Frameworks like AutoGen, CrewAI, and LangGraph each give you the plumbing, checkpoints, persistence, message-passing primitives. None of them decide your governance for you. Scope, policy, ingestion, and evaluation are still yours.

“In multi-agent memory, sharing buys collaboration, and grows the attack surface.”

Shipping it

In production, memory is a system.

Adding memory to a real product is not “install a vector DB.” You need an API surface, separate read and write paths, background extractors, multi-tenant isolation, latency budgets, observability, and a plan for the day you change your embedding model. The list is long because production is long.

Work in progress module

This section is an active draft. The architecture diagram and latency sliders are teaching sketches with toy defaults; they are not a complete or production-validated implementation guide yet.

Click any component to see its role.

Two services, sharply separated: the agent runtime on the request path, and the memory service as a side-quest. Background workers do the slow, expensive work - extraction, summarization, re-embedding, decay, without blocking the user.

Multi-tenancy

Three ways to isolate.

Pros

Cons

Best fit

Most production systems with mixed tenant sizes.

Small tenants share a collection; enterprise tenants get dedicated.

The other axis is time. Memory retrieval runs on eligible turns, so every millisecond on that path shows up in p95. The interesting optimizations aren't the obvious ones, they're things like skipping retrieval entirely on queries that don't need it, caching the top-N facts per user in a KV, and running the dense, sparse and graph retrievers in parallel rather than sequentially.

Latency budget

Where the p95 goes.

Drag the sliders. The bar fills with each stage's contribution. Target is 800 ms, anything past that is felt.

Storage tiers

Hot, warm, cold.

Not every memory needs to be a vector lookup. A KV cache for the top facts pays for itself in latency and tokens.

Active user / project facts. KV or in-memory cache.

Recent episodes, preferences. Vector + full-text index.

Raw event log, archived sessions. Object storage.

The minimum surface area you can ship and still call it a memory system. Everything else (consolidation, supersession, analytics) is built on top of these three.

POST/memory/events

Append raw events. The write queue handles dedup and indexing asynchronously.

POST/memory/search

Hybrid retrieval with scope, time, and top_k. Server enforces filters from the JWT.

DELETE/memory/{id}

Forget. Propagates to derived indexes. Deletion lineage is its own audit event.

None of this is exotic, it's the same architecture you'd build for any retrieval product. The lesson is just that agent memory is a retrieval product, not a feature flag. Treat it like one and the rest follows.

Capstone

Build your own agent memory.

Everything in this essay collapses into a small set of knobs. Choose a backend, set the limits, flip on the techniques you trust. Send a message and watch each stage of the pipeline fire, retrieval, generation, governance, write.

Memory backend

Embed everything, retrieve top-k by cosine similarity.

Context window

Retrieval k

Send a message to begin. Try one of the suggestions below.

Try this: send the same question with the governance gate off, then turn it on and send a contradicting fact. The inspector will show old entries marked supersededinstead of being overwritten, and the credit-card pattern never makes it past the gate.