Lab brief L-04AI memory · agentsStatus — Research

Lethe.

A governed memory layer for AI agents, built on the principle that forgetting is a feature, not a failure.

Role — Architect & builderStack — Vector store · consolidation engine · provenance graphStage — Research prototype

§ 01The idea

The upper floors of machine memory.

Retrieval-augmented generation is recall on demand: the ground floor of cognition. It fetches what's relevant now. But the agents we're being asked to build are expected to carry a case across weeks, notice that today's message contradicts last month's, and learn what is worth keeping. That isn't retrieval. It's memory, and memory is a building with several floors.

Lethe is an experiment in the upper floors: a layer that sits above retrieval and decides what an agent should retain verbatim, abstract into a summary, or let decay, modelled loosely on how biological consolidation promotes a few experiences and lets the rest fade. Every durable memory carries its provenance, so a memory you can't trace is a memory the system won't trust.

The name is deliberate. In a system that hoards everything, relevance drowns in volume and the compliance surface grows with every byte you refuse to forget. Lethe treats principled forgetting as a first-class design goal.

§ 02Why it's different

Four moves that turn retrieval into memory.

Separate the floors. Retrieval, working context, and consolidated memory are distinct stores with distinct lifetimes, so stale facts don't leak into live reasoning.
Write the forgetting policy first. Decide what decays, what is abstracted into a summary, and what is kept verbatim, before deciding what to store at all.
Make memory auditable. Every durable memory records where it came from, when, and on whose authority. Provenance is structural, not bolted on.
Govern the upper floors hardest. Recall on demand is low-stakes; consolidated, institution-level memory is where bias, drift, and liability accumulate. The controls live there.

§ 03How it works

Capture, consolidate, forget, prove.

Working memory

01 · Capture

The live thread of a task is held separately from anything durable: what we're doing, and why, right now.

The replay pass

02 · Consolidate

A scheduled pass promotes the few interactions that matter into durable memory and abstracts the rest into summaries.

The decay policy

03 · Forget

An explicit, per-class schedule governs what fades, when: the same rigour as a data-retention policy, because it is one.

Provenance graph

04 · Prove

Every durable memory links back to its source, time, and authority, so any recalled fact can be traced and defended.

"A memory that keeps everything has no opinion about what matters. Neither does a system."

From the dispatch: The Cartography of Memory

§ 04What I'm testing

Open questions

Lethe is research, not a product: the most experimental thing in the Lab. The questions I'm chasing: what a good default forgetting policy even looks like across task types; whether consolidation can run cheaply enough to be continuous; how to keep provenance intact through abstraction and summarisation; and whether "principled forgetting" measurably improves an agent's reasoning over a hoard-everything baseline.

If you're working on agent memory, evaluation, or governance, I'd genuinely like to compare notes.

Retrieval is the ground floor. This is the building.

Compare notes → ← More experiments

An exploratory research prototype by Dr. Nabeel A. Khan.

Fin · Sheet 10