Memory System

Lamis said this memory system approach is not only applicable to programming tasks—he himself uses the memory system to capture his writing preferences and slide style for presentations.

The grounding mechanism anchors every fact in a workflow to its original source, while the memory system continuously retains enterprise background information; the combination of the two is the key support for reliable execution of long-horizon tasks.

A complete memory system generally needs to include write filtering of memories, time-based decay or version overwriting, and compression and summarization after recall.

Context window length cannot be equated with memory capability; the window determines how much content can be held at one time, while a memory system addresses which content is selected to enter the limited window.

Practical advice for labeling fallback paths: if any client path may store locally, its response must state so in the result, not in stderr.

Practical advice for ordering destructive operations: validate before deleting, and specifically test malformed-input paths.

Memory systems should output normalized evidence packets with source identifiers rather than raw fragments, so that downstream modules can perform traceable filtering and combination.

End-to-end validation means testing along the complete chain from user input to the final answer, rather than only verifying whether a single component throws an error; for a memory system one can design probe cases such as asking about the same fact in three different wordings, switching between topics A/B, having two users ask alternately, and checking that new topics do not trigger fabrication.

Call records should be kept in their own database, separate from the notes the bot recalls, because if written into the same memory, every recorded answer would become a candidate in the next recall, including the bad answers.

Most current mainstream LLM agent frameworks emphasize persistent memory and context accumulation, treating history length as a sign of capability.

According to the ForgetEval research, when leading memory systems were scored on test cases written by people other than the authors themselves, internal test scores of 63%-68% plummeted to 28%-33%.

For AI coding agents, the memory system may become a more important differentiator than model capability itself, because model capabilities are rapidly converging, while accumulated project context is a truly hard-to-replicate asset.

The most common deviation in the first week or two of using a memory system is over-writing, saving a large amount of temporary information as rules as well, which degrades recall quality; afterward it is easy to swing to the other extreme of recording nothing and falling back to the old path of repeated explanations.

A memory system is usually accomplished through the collaboration of three components: writing (extracting facts, rules and decisions worth retaining long-term from conversations), retrieval (recalling relevant entries by semantic similarity or keywords), and injection (automatically splicing the recalled content into the system prompt or the first round of context).

Rule rot is the most common degradation pattern in long-term use of a memory system; its cause is that the cost of writing is low while the cost of cleaning up is high, and as entries increase, the recall probability distribution during retrieval is diluted.

An actionable practice to combat rule rot is to do a rule audit once a month: print out the current rule set and read it through, merge synonymous entries, delete invalid entries, and split overly general entries into decidable, specific entries.

The memory system injects long-term rules and the current task through separate channels, structurally alleviating the dual losses of the context window being filled up and lost in the middle.

Even if a model can retrieve relevant memory context, it does not necessarily truly constrain its own behavior during generation based on that context (writing to memory does not equal following memory)

In the composite score for memory retrieval, similarity has the largest weight, recency accounts for a small proportion, and the result is multiplied by an importance factor with a range of approximately 0.8 to 1.2

Memory systems and RAG systems are fundamentally both data retrieval problems, where code finds the right data and places it into the prompt