Deep Dives
Architecture, engineering, and benchmarks
Per-User Memory for AI Products: Multi-Tenant Patterns
How to give every user isolated memory in a multi-tenant AI product: hard boundaries vs soft partitions, testing for cross-tenant leakage, GDPR deletion, and scaling.
The 10 Things to Look For in an Agent-Memory System
The 10 things that separate a real agent-memory system from a demo: retrieval, conflict resolution, freshness, security, self-hosting, observability, and more.
8 Agent Memory Use Cases (With Real Examples)
The real agent memory use cases teams build on Hindsight, from coding agents to customer-facing products to automations, each with a concrete example.
recall vs reflect: Search Your Agent's Memory, or Ask It
Hindsight gives your agent two ways to read its memory. recall retrieves the relevant facts (no LLM, instant). reflect reasons across them to synthesize an answer. When to use each.
Your 1M-Token Context Window Is Not Memory
A context window is working memory: it forgets everything when the session ends and degrades long before it fills. Why a bigger window is still not agent memory.
One Bank or Many? A Field Guide to Structuring Agent Memory
Every Hindsight integration tells you to set a bank_id, but not how to decide what a bank should be. A bank is a recall boundary: how to scope agent memory, and when to reach for tags instead.
Inside retain(): What Actually Happens When Your Agent Remembers
Calling retain() is one line of code. Underneath, a sentence becomes extracted facts, resolved entities, a knowledge graph, and consolidated observations. Here is the whole write path.
Entity Resolution in Agent Memory: One Person, Many Names
How agent memory decides that Sarah, Sarah Chen, and she are one person, and why Hindsight resolves entities with a co-occurrence graph, not embeddings.
Staleness-Aware Memory: When Your Agent Should Verify Before It Trusts
Hindsight's consolidated memory (observations and mental models) is derived from raw facts, so it can lag behind new writes. Instead of silently serving a stale snapshot, Hindsight flags how far behind each layer is, and the reflect agent verifies against raw facts before trusting it.
How Hindsight Learns: A Deep Dive Into Mental Models
Mental models are the top tier of Hindsight's learning hierarchy — persistent, self-refreshing understandings of the topics your agent asks about every session. A walkthrough from the schema up.
The Consolidation Problem in Agent Memory
Agents that remember everything remember nothing useful. A four-lever framework — importance, merge, decay, eviction — for consolidating agent memory.
The Case Against External Vector DBs for Agent Memory
Most 'add memory to your agent' tutorials start with installing a vector database. They shouldn't. Why vector DBs are the wrong default for agent memory.
How Hindsight Scales
A design analysis of how Hindsight's memory operations scale with data volume — what costs grow, what stays bounded, and why.
The Missing Layer in Every Agent Harness
Modern agent harnesses ship with tools, MCP, and IDE integrations — but no memory. Why that's the missing layer, and how harnesses are starting to fix it.
Your Agent Is Not Forgetful. It Was Never Given a Memory.
Why agents seem forgetful, and why memory is different from context windows and retrieval. How Hindsight adds long-term memory to agents.
How We Built a 4-Way Hybrid Search System That Actually Runs in Parallel
Sequential async queries were killing our retrieval latency. Here's how we built a true 4-way parallel hybrid search system with asyncio and RRF fusion — then evolved it further with connection sharing, cross-encoder reranking, and multiplicative boost scoring.
Agent Memory Benchmark: A Manifesto
Agent Memory Benchmark: A Manifesto
How We Built Disposition-Aware Agents That Actually Think Differently
Hindsight — Disposition-Aware Agents
How We Built Time-Aware Spreading Activation for Memory Graphs
Hindsight — Stories, Not Rows
How We Solved Memory Conflicts in Hindsight
Learn how Hindsight handles contradictory information by tracking temporal evolution and preserving history in its memory consolidation system.