P

Loading...

Memory Layer โœ๏ธ
Tech

Memory Layer โœ๏ธ

October 24, 2026by PETER
133
by PETER

Intro

  • This is the ๐Œ๐ž๐ฆ๐จ๐ซ๐ฒ ๐‹๐š๐ฒ๐ž๐ซ!

  • ๐Ÿ‘จโ€๐Ÿซ What does that mean? These are tools that let your computer answer, "Of course!" when you ask, "Hey, you remember what I said last time, right?" even after turning it off and back on.

  • They retain memories across sessions, work across various harnesses (Cross-Harness) such as Claude Code-Codex-Cursor, and intelligently retrieve only the memories they need. And best of all, every single one is โœจโœจ๐Ž๐ฉ๐ž๐ง๐’๐จ๐ฎ๐ซ๐œ๐žโœจโœจ


1. ๐’๐ฎ๐ฉ๐ž๐ซ๐ฆ๐ž๐ฆ๐จ๐ซ๐ฒ ๐Ÿ’ช

  • A memory API that knows "who this person is." It seems great for building AI Powered Apps.
    https://github.com/supermemoryai/supermemory

  • Its key feature is that it automatically manages 'user profiles.' So rather than coding work history, it focuses on 'people and conversation personalization.' Its architecture handles memory + RAG together with a single hybrid query!

    image

  • Want it to truly understand you through a knowledge graph? Supermemory is the answer.๐Ÿ˜Ž


2. ๐‡๐ข๐ง๐๐ฌ๐ข๐ ๐ก๐ญ ๐Ÿค”

  • Memory that doesn't just remember, but "learns." It turns memories into understanding.
    https://github.com/vectorize-io/hindsight

  • It manages memory through a total of four stages: โ‘  What is true? โ†’ โ‘ก What have I experienced? โ†’ โ‘ข What did that lead me to believe? โ†’ โ‘ฃ So what should I do now? (Mental Model)

image

  • Stages 1 and 2 contain objective records, while stages 3 and 4 hold insights it has developed and learned on its own, which the model then uses to respond. Based on this, it continuously keeps Knowledge Pages up to date...! Isn't that an amazing technology?(How is this free)

3.๐š๐ข-๐ฆ๐ž๐ฆ๐จ๐ซ๐ฒ ๐Ÿง 

  • File-based shared memory specialized for handoffs between agents
    https://github.com/akitaonrails/ai-memory

  • Rather than general-purpose knowledge, it specializes in remembering "this Repo." It seems well suited to situations like qm, where each Slack channel has a separate workspace and anyone coming in needs to be able to take over.

  • It also allows unused memories to gradually fade and frequently used ones to remain longer, even without an LLM. It does a good job of removing meaningless memories using a survival score formula and DBSCAN. Below is the memory aging formula. โœ๏ธ

    image

  • But it's somewhat cumbersome because you have to run the server yourself. So it seems like a good choice when multi-user logs, audit logs, and similar features are important within a team or company that shares machines.


4.๐๐ž๐š๐œ๐จ๐ง โš™๏ธ

  • It selects useful items from records and turns them into knowledge. ๐Ÿช„
    https://github.com/Asymptote-Labs/agent-beacon

  • It combines not only prompts and responses, but also tools, commands, file changes, approvals, MCP, and tokens into a single Trace. It identifies successful workflows, corrections, and debugging patterns, then reuses only those approved by a person! Beacon has this concept of "promoting memories into skills."

  • Thanks to these features, Beacon lets you inspect Traces of everything happening across all agents in real time. ๐Ÿ‘€ Perhaps because of that, it also integrates well with security and operations, making it useful when you need to track and audit what agents actually did.

    image


5. ๐œ๐ฅ๐š๐ฎ๐๐ž-๐ฆ๐ž๐ฆ ๐Ÿ‘พ

  • The Memory Layer veteran, Claude-mem.
    https://github.com/thedotmack/claude-mem

  • It automatically observes and compresses coding sessions, then injects them into the next session. It's the most popular choice. If you're thinking, "I don't care about the complicated stuff! I just need it to remember well!" then simply install this and you're done! It's only called claude-memโ€”you can install it for Codex, Cursor, and more.

    image


Summary

Memory focusStorage formatLLM required?Primary users
supermemoryPeople (profiles)Graph engine + embeddingsYesAI app developers
HindsightLearning (beliefsยทmodels)Postgres + pgvectorYes (when storing)Agent developers
ai-memoryRepo + handoffsgit Markdown wikiNo (by default)Developersยทteams using multiple CLIs
BeaconComplete session recordsJSONL (OTel)When extracting knowledgeTeamsยทsecurity personnel
claude-memCoding observationSQLite + ChromaYes (when compressing)Individual Claude Code users

Closing ๐Ÿณ

  • It's 2026, and Agents are getting increasingly Fancy with their memories too! ๐Ÿค—โœจ

โœจHAPPY BUILDING!๐Ÿ‘ทโ€โ™‚๏ธ
4
Buy me a coffee

0 comments

Please login to leave a comment.