Most RAG setups fail because they treat memory like a static filing cabinet. When every transient bug fix or abandoned rule is stored forever, the context window eventually chokes on noise, spiking token costs and degrading the agent's reasoning.

This implementation experiments with a biological approach by using the Ebbinghaus forgetting curve to manage context as a living substrate. Memories are assigned a "strength" score where each recall reinforces the data and flattens its decay curve (spaced repetition), while unused data eventually hits a threshold and is pruned.

To solve the "logical neighbor" problem where semantic search misses relevant but non-similar nodes, a graph layer is layered over the vector store. Benchmarked against the LoCoMo dataset, this reached 52% Recall@5, nearly double the accuracy of stateless vector stores, while cutting token waste by roughly 84%.

Built as a local first MCP server using DuckDB, the hypothesis is that for agents handling long-running projects, "what to forget" is just as critical as "what to remember." I'd be interested to hear if others are exploring non-linear decay or similar biological constraints for context management.

GitHub: https://github.com/sachitrafa/cognitive-ai-memory

I know everybody seems to want the agent to remember every conversation they've ever had with it, but I just don't see the value in that. In fact, it seems to hurt productivity to have the agent second guessing me based on something I said yesterday. Every time I've used any memory system, the agent gets distracted from the current tasks based on previous conversations and branches of development...often comingling unrelated projects (I work on code for work, open source projects, a bunch of unrelated side projects, etc.) and trying to satisfy requirements that don't make sense.

I've stopped trying to achieve general "memory". I just ask the agent to thoroughly, but concisely, document each project. If it writes developer documentation and a development plan/roadmap, as though a person was going to have to get up to speed and start working on the project, it provides all the information the agent needs tomorrow or next week to pick up where we left off.

The agent is not my friend. I don't need it to remember my birthday or the nasty thing I said about React last week. I need it to document what anyone, agent or human, would need to know to get productive in a particular repo, with no previous knowledge of the project.

Good, concise, developer and user documentation and a plan with checklists solves every problem people seem to think "memory" will solve: It tells the agent what tech stack to use (we hashed it out in planning), it tells it what commands it needs to run and test the app, it covers the static analysis tools in use (which formalizes code style, etc. in a way a vague comment I made a month ago cannot), and it is cheap. Markdown files are the native tongue of agents. No MCP, no skills, no API needed. Just read the file. It works for any agent, any model, and any human just getting started with the project.

Basically, I think memory makes agents dumber and less useful. I want it to focus on the task at hand.

It strikes me as funny how we want to get super AI inteligence but keep trying to anthropomorphizing all AI aspects to make it more "human". IMHO, if we keep doing it we will create Human AI with all errors and deficiencies humans have.

I planned and supervised the build of an ambient recall system, where a 4b model looks at the last 3k or so of context and picks through the RAG database for high ranking memories to inject, as well as mineable things to mark. Injections happens about 1/5 turns on most technical topics, data picked from prior design docs and data sheets mostly. At session wrapup the inference model goes back and rates all the memory injections in a frontmatter section, then looks at all the memory suggestions to commit those it finds memorable to the RAG database. Manual memorisation and RAG search are also available inline in the chat to both the user and the model. It also allows the main model to spawn little models as minions to work on repetitive simple tasks.

Seems to maybe be useful but I’m not sure yet.

It's the cumulative weighting based on the softmax output? Is it per layer?

I haven’t had much like with memory implementations. I tried a few.

What I do now is preserve all my claude code conversations and set the context from there.

This allows me to curate memory and it’s been the best way so far.

I am also building a similar memory structure and decay mechanism for my local agent project, where I also use Ebbinghaus. One of the challenge I face is how to decide effectively what to save in the memory: Is it the model to decide what is important, summarize and save it to the memory? How to avoid redundancy and categorize the memory correctly so you could get the right hit and decide what to forget. I would love to learn more about your approach and what your thoughts on those points

I only use a decay function to see how "hot" a chunk is - not for forgetting old ones. What concerns me more are memory chunks with errors in them - they need to be corrected/removed by some other mechanism, not by decay (since they might get retrieved often).

you just make Alzheimer a feature lol , but seriously this is very interesting

Good perspective on this.

on the other "biological memory" post in so many weeks, I pointed out that the decay rate shouldn't be based on a real clock but a lifetime of it's use within the coding session. Elsewise your memory fades even when there's no process change (eg, coder goes on vacation). I'm not going to check whether thats true here, but it seems like a naive first assumption thats failed conceptualization.

The other comment is that spatial memory is probably a better trigger for memory, so if you're not tracking where the coding session starts, the folders it's visits, etc, then you're not really providing a good associative footpath for the assistant to retrieve whats important for any given project.

I am sorry but the whole "biological memory" thing seems like marketing fluff on basic cache mechanisms.

You said it cuts token usage by 84% but isn't that typical for any typical chunked RAG system?

And why did you specifically chose to test against the LoMoCo dataset when there's a lot of issues with it and it being very easy to cheat?