there’s a subtle issue i keep running into when building agent systems

people talk about memory like it’s solved because they added a vector db

but in practice, the system still forgets decisions, reintroduces context, and behaves inconsistently across sessions

so the real problem isn’t storage

it’s structure + retrieval reliability over time

What i changed in my setup instead of trying to “store more context”

i rebuilt memory as a layered system that separates capture, compression, structure, and correction

architecture overview 1. capture layer (raw persistence) everything is logged first without filtering

daily files only

goal is simple: never lose information at ingestion time

2. distillation layer (information compression) a scheduled process (cron-based) converts raw logs into stable memory

only long-term relevant data is kept:

persistent preferences

decisions

stable facts

active projects

this is where noise gets removed

3. atomic memory structure memory is split into single-concept files

no mixed documents

tools

people

projects

ideas

this improves retrieval consistency significantly

4. implicit graph structure instead of using a graph database

files explicitly reference related files using markdown links

this creates a lightweight semantic network without extra infrastructure

5. retrieval optimization layer this is where most systems fail in practice

instead of relying purely on embeddings, i enforced:

synonym expansion (fr/en)

multiple semantic formulations per concept

keyword redundancy

rephrasing of key ideas in different contexts

this reduces retrieval blind spots caused by embedding mismatch

6. self-improvement loop retrieval failures are logged and periodically reviewed

the system adjusts:

file structure

keyword sets

placement of information

missing links between concepts

over time, memory quality improves instead of degrading

why this approach most systems optimize for retrieval accuracy in isolation

but memory in agents is not just retrieval

it’s also:

consistency over time

stability of decisions

ability to re-use context without re-injection

so the focus shifted from “better embeddings” to “better information architecture”

outcome so far after running this structure for a while:

fewer repeated context injections

more consistent behavior across sessions

reduced token usage due to better reuse of stored context

fewer contradictions in tool usage and decisions

the model didn’t fundamentally change

the system around it did

open question plz i’m still exploring:

how much distillation is optimal before losing nuance

whether explicit graph modeling would outperform implicit linking

how redundancy in retrieval scales in larger memory graphs

curious if anyone has pushed this further in production agent systems :))

[留言]

为什么值得关注

能改变理解方式，而不只是重复常识；符合当前抓取需求；它提供了新的理解或解释，而不只是表面观点

来源：reddit，领域：tech，保留分：0.58