Agent memory is a control plane, not a vector replacement

TencentDB Agent Memory showed 29.1k GitHub stars, 2.8k forks and 63 commits when I inspected its current repository on 14 October 2026. A viral X post says the project has killed the vector-database industry.

The primary architecture says the opposite. TencentDB Agent Memory still combines BM25, vector retrieval and reciprocal-rank fusion. Its useful addition is the control plane around retrieval: versioned memory assets, ownership, access rules and a declared agent loadout.

TencentDB Agent Memory still uses hybrid retrieval

The current TencentCloud repository describes four asset classes: Chat Memory, Skills, Wiki and CodeGraph. Long-term recall normally starts from higher-level persona or scenario layers, then falls back to lower-level evidence through BM25 plus vector retrieval plus RRF.

That is not a vector replacement. It is a system deciding what may be retrieved, which version is valid, who may see it and which agent should receive it.

Layer	Job in the current design
Retrieval	BM25, vectors and RRF find candidate evidence
Memory assets	Chat Memory, Skills, Wiki and CodeGraph package reusable context
Control plane	Ownership, version, status, visibility, ACLs and agent bindings govern use

The architectural distinction is retrieval versus governed reuse, not vectors versus no vectors.

My pilot would test governance before recall quality

This is labelled editorial judgment; I did not install the multi-service stack or run it on a Trigent or client system. For a hypothetical enterprise pilot, I would first create one private memory, one team memory and one revoked memory, then verify which agent receives each version and whether the UI can drill back to the source.

Only after that would I score recall relevance. A memory system that finds the right fact for the wrong agent is not more capable; it is a permissions failure with good retrieval.

The current benchmark is useful but narrow

The default branch reports PersonaMem accuracy rising from 48% to 76%. I could not independently verify that result. The older 61.38% OpenClaw token-saving figure repeated in the viral post is absent from the current default-branch benchmark table.

The README also says private-repository support is still being refined, automated memory routing remains under iteration, and Wiki/CodeGraph builds are asynchronous. Those are reasonable beta limits. They are also why the project deserves an acceptance test, not an industry obituary.

What's in it for you

• Keep your vector-store evaluation; add memory ownership, versioning, revocation and source drill-down to it.
• Judge an agent-memory layer by governed reuse across sessions, not by a retrieval screenshot alone.

Agent memory does not replace retrieval; it decides which retrieved evidence may become reusable operating context.

Sources

• Divyanshi — TencentDB Agent Memory X post, 12 October 2026
• TencentCloud — TencentDB Agent Memory repository, accessed 14 October 2026