cavendish
Standing answerTestedstrength · strong

Which memory layer should a new agent use?

Under ten sessions of history: none, just the context window. Over ten: the lab's episodic store with summarised recall via the memory adapter. Not a vendor memory product yet.

Tier is not strength

Tier says how much we know. Strength says how hard we are telling you to act. Scored independently.

Evidence tierTested
Strengthstrong
OwnerTOTom Okafor
Last validated21 Aug 2026
Review by5 Oct 2026
Half-life45 days
Citations31
Asked96× this quarter
VerticalsCross-sector, Banking, Insurance
decay32d until review

Machine-readable target

{
  "taskType": "multi-session-agent",
  "configKey": "agent.memory.store"
}

Nothing consumes it yet. Day two: findings ship as defaults into the gateway, routing config and skill library.

Body

As of 21 August 2026: if the agent will hold fewer than ten sessions of history, do not add a memory layer; the context window and the system of record are enough. Above that, use the structured episodic store with summarised recall through the memory adapter (r-memory-layer). Do not adopt either of the two vendor memory layers launched in August as the default; neither documents its consolidation policy, and that is the part that carries the value.

Evidence: x-memory-bench on a 40-session support corpus. Episodic plus summarised recall won on precision by 19 points and on latency by 0.6× above 50k tokens of history (c-agentic-memory-1, c-agentic-memory-2). Graph memory is a special case for entity-heavy workloads and otherwise degrades (c-agentic-memory-3). The vector-store design is in the graveyard as g-vector-memory-v1.

Caveat: no PII in the episodic store until a forgetting mechanism that satisfies APP 11 exists. The two vendor layers are on the Q4 re-bench; this answer may change after that.

TOSigned Tom Okafor · Research engineer · agents · 21 Aug 2026

What it rests on

Retrieval latency dominates memory-layer cost above ~50k tokens of accumulated history, regardless of store type.

Tested c-agentic-memory-1
84%

Summarised episodic recall beats raw chunk retrieval on decision-consistency tasks by a wide margin (>15 points).

Tested c-agentic-memory-2
79%

Graph-structured memory improves entity-heavy tasks and degrades general tasks; it is not a default.

Tested c-agentic-memory-3
66%

Field

Agentic Memory System

Experiment · validated

Memory layer bake-off on a 40-session support corpus

Graveyard · superseded

Vector-store memory as the agent's long-term memory Remembered everything, retrieved nothing.