Memory for AI Agents

Memory for
AI agents

How an AI assistant remembers you: the plan, and eleven short reads. A card file with two dates, a secretary who never tears a page out, and a librarian with three ways to search.

13 pieces ~49 min in total Plain English
Start here
The whole idea

In plain English

The problem, the card file, the secretary and the librarian, and why every card carries two dates. No jargon.

5 min read
The build

The plan

What we take from Mem0, HydraDB and the papers, the data model, the module layout, benchmark targets and milestones.

8 min read
The eleven articles
01

Why do AI agents forget?

An LLM is a genius with amnesia. History grows, memory stays flat.

3 min read
→
02

What kinds of memory does an agent need?

The sticky note and the card file: semantic, episodic, procedural.

3 min read
→
03

How does an agent decide what to remember?

The secretary: extract, dedup, and never tear a page out.

3 min read
→
04

How does an agent find the right memory?

The librarian: meaning, words and names, then recency and importance.

3 min read
→
05

How does memory stay true when life changes?

Two clocks on every fact. Git for facts.

3 min read
→
06

Why does a memory system need a delete button?

The hoarder's garage: dedup, supersede, decay, erase.

4 min read
→
07

How does Mem0 work?

The 2026 rewrite: ADD-only, rich memories, entities, three-signal search.

4 min read
→
08

How does HydraDB think about memory?

Memory is a database problem: append-only, bitemporal, provenance.

4 min read
→
09

How is agent memory measured?

LoCoMo, LongMemEval, BEAM, and why a score needs its token cost.

3 min read
→
10

What are we building?

The hybrid, one fact walked end to end, and the code layout.

3 min read
→
11

How does memory scale to a million users?

Async writes, shard by user, shrink the vectors with TurboQuant.

3 min read
→