Interface: IWorkingMemorySummaryBuffer

Defined in: internal/working-memory/types.ts:271

A summary + facts kept by the cheap model after every updateEveryTurns turns, plus the last recentTurns turns verbatim.

Read by WorkingMemory.load and saveContext.

Properties

maxFacts?

optional maxFacts?: number

Defined in: internal/working-memory/types.ts:301

Cap for the number of facts.

Default

30

---

maxSummaryChars?

optional maxSummaryChars?: number

Defined in: internal/working-memory/types.ts:295

Cap for the summary, in characters.

Default

1500

---

recentTurns?

optional recentTurns?: number

Defined in: internal/working-memory/types.ts:283

Verbatim turns sent after the summary, newest last.

The tail the summary does not cover yet.

Example

20 - a chat of normal length fits a 128k model with room for the knowledge-base chunks

Default

20

---

type

type: SUMMARY_BUFFER

Defined in: internal/working-memory/types.ts:273

The discriminator of TWorkingMemoryStrategy: this one is the summary buffer.

---

updateEveryTurns?

optional updateEveryTurns?: number

Defined in: internal/working-memory/types.ts:289

Run the LLM update once the thread grew by N turns (IThread.turnCount) since the last update.

Default

1