Long-term memory
Memory that outlives the session: writing it, consolidating it, and — the part everyone forgets — letting it decay. MemGPT, memory streams, and why forgetting is a feature.
A session ends and the context window is wiped. If the agent should remember the user next time, that memory has to be written to a durable store, retrieved later, and eventually forgotten. Long- term memory turns a stateless model into something that accumulates experience — and inherits all the problems of a system that accumulates anything.
Start here
Long-term memory is a loop, not a bucket: observe something worth keeping, write it, recall it when relevant, and let the unimportant parts fade. Forgetting is not a bug to fix — an agent that remembers everything retrieves noise, pays for it in context, and never develops judgement about what matters.
Gradient strength reaching each earlier time step (right = last step, left = far past).
usable memory horizon ≈ >60 steps
The gradient is below 1% of its strength after the shaded span (63 steps back). Nothing earlier than that is being influenced by what arrived — the mechanism behind untrainable long-range dependencies in plain RNNs.
|factor|^(distance back) on a log scale: the multiplicative anatomy when the same recurrence weights repeat at every unrolled step. Real curves wobble with W_h’s spectrum and gate states, but every honest long-RNN shows a usable horizon of tens of steps — this is why plain RNNs can’t learn long-range dependencies.
Watch stored items lose relevance over time unless they are reinforced.
Two reference designs
- MemGPT treats the LLM like an operating system: a small "main context" (the window) and a large external "archival" store, with the model paging memory in and out and summarising when the window fills.
- Generative Agents keep a memory stream of observations, score each by recency, importance, and relevance, and retrieve the top items — with periodic reflection that synthesises higher-level summaries from the raw stream.
Consolidate and forget
- Consolidation — summarise many raw events into fewer, higher-level memories, the way reflection does. This is compaction applied to a store instead of a transcript.
- Decay — let importance fade with time unless reinforced, so the store stays a signal of what mattered, not a landfill.
- Correction — memory must be editable. A wrong memory is trusted precisely because it is remembered.
Careful
Long-term memory is where privacy, staleness, and unbounded growth concentrate. Storing user data raises consent and deletion obligations; an un-editable memory means a mistake persists forever; an un-forgetting store grows until retrieval is useless. Decide what deserves to be remembered and for how long before you build the write path.
Check yourself
Eduspheria wiki · Agentic AI, Memory & context
0 / 4 answered
Next: the engineering of what actually sits in the window.