Is a vector database enough for AI agent memory?

For recalling related notes, often yes. For memory an agent acts on, usually not: similarity search returns the closest text, not the current truth, so an outdated or contradicting note can outrank the right one. Agents also need the present value of a fact, conflict detection, and provenance, which a vector index does not track.

What is the difference between RAG and agent memory?

RAG retrieves passages from a mostly static corpus to answer a question. Agent memory is written by the agent while it works: facts change, get superseded, and can contradict each other, so the store has to track state over time, not just similarity.

How do multiple agents share memory?

Give them one shared store keyed by the task or team, so every agent writes to and recalls from the same place. The store should flag conflicting claims instead of letting the last write win silently, and record which statement each fact came from.

Does SynapCores agent memory need an LLM?

No. REMEMBER, RECALL, CURRENT and TRACE run without a chat model: facts are extracted with built-in rules, and an LLM, if available, improves extraction later during consolidation.

AI Agent Memory: Why a Vector Database Isn't Enough (2026)

Most "give your agent memory" tutorials do the same thing: embed every message, store the vectors, and at the next turn retrieve the top five most similar chunks. It works in the demo. Then the agent tells a customer the old price, or two agents in a crew act on contradicting notes, and nobody can say where the bad fact came from.

The problem isn't the vector database. Similarity search answers "what text is close to this question?" An agent acting on memory needs answers to three other questions:

What is true now? The plan is Pro, not the Free plan the user mentioned in March.
Do my sources disagree? The researcher says the price is $19, the planner says $29.
Why do I believe this? Which statement, when, and what did it replace?

A vector index stores none of that. A memory system has to.

Where vector-only memory breaks

Stale facts win on similarity. "I'm on the Free plan" and "I upgraded to Pro" are nearly the same sentence to an embedding model. Ask "what plan is the user on?" and either can come back first, or both, and the model guesses.

Contradictions are silent. If two agents post conflicting values, a vector store keeps both rows. Retrieval returns whichever is closer to the query wording. Nothing tells the agent there is a disagreement to resolve.

No provenance. When a fact turns out to be wrong, you want the original statement, the time it was recorded and the method that extracted it, so you can fix the source rather than delete rows by hand.

Duplicates pile up. Agents repeat themselves. Without consolidation, the same fact is stored ten times and crowds better context out of the prompt.

What agent memory needs on top of vectors

Need	Vector store	Agent memory
Recall related notes by meaning	✅	✅
Current value of a fact	❌ (closest text wins)	✅ authoritative state
Detect conflicting claims	❌	✅ flagged, with both claims
Provenance (source, time, method)	❌	✅
Supersede old facts over time	❌	✅ temporal validity
Consolidate duplicates	❌ (manual)	✅

Vectors stay inside the memory system. They are the recall layer, not the whole memory.

The same idea in five statements

SynapCores has agent memory built into the database as a first-class object. Here is a shared board for a team of agents working on a product launch. Everything below runs on the free Community Edition 2.2.0, with no chat model required:

-- One shared memory, partitioned by task. Episodes, facts, current state,
-- provenance and relationships are created for you.
CREATE MEMORY team_board IDENTITY task_id;

-- Agents write what they learn.
REMEMBER team_board FOR task_id = 'launch' 'The researcher found that EMEA buyers prefer monthly billing.';
REMEMBER team_board FOR task_id = 'launch' 'The launch price is 19 dollars per month.';
REMEMBER team_board FOR task_id = 'launch' 'The launch price is 29 dollars per month.';

The third REMEMBER doesn't silently overwrite the second. It comes back with "status": "conflict", because two claims of equal authority now disagree about launch_price.

-- Recall by meaning: current state first, then related facts and episodes.
RECALL team_board FOR task_id = 'launch' ABOUT 'pricing';

-- Ask for the authoritative value of one attribute.
CURRENT team_board FOR task_id = 'launch' ATTRIBUTE launch_price;

CURRENT doesn't guess. It returns status: conflict with both competing claims (19 dollars per month and 29 dollars per month), their authority, confidence and timestamps, and the reason it couldn't pick one. That's the moment a coordinator agent or a human should step in, instead of shipping the wrong price.

-- Why do we believe the value we have?
TRACE team_board FOR task_id = 'launch' ATTRIBUTE launch_price;

TRACE returns the original statement verbatim, when it was recorded, the extraction method, and what the value superseded, if anything.

Using it from an agent framework

The memory lives in the database, so any agent that can run SQL or call an MCP tool can use it: OpenClaw, Claude Code, LangGraph, or a voice agent. For OpenClaw there's a ready plugin: @synapcores/openclaw-memory detects the memory object and maps it to OpenClaw's memory tools, including memory_current (the authoritative value, with an explicit conflict outcome) and memory_trace (why a value is believed).

If you'd rather control the schema yourself, the lower-level MEMORY_STORE / MEMORY_RECALL / MEMORY_UPSERT functions and plain tables with EMBED() still work. The multi-agent blackboard recipe builds one by hand, and the self-consistent memory recipe gates writes with a contradiction check.

When a vector database is enough

Be honest about the simple case. If your "memory" is a fixed set of documents and the agent only reads it, that is RAG, and a vector index, ideally with a relevance floor, is enough. Reach for agent memory when the agent writes facts that change over time, or when several agents share what they know.

Try it

Download the free Community Edition for macOS, Linux or Docker.
Paste the five statements above into the SQL editor.
Add a third agent's claim and watch CURRENT flag the conflict.