A six-part series on why AI coding agents forget everything, why standard vector RAG is not enough, and what it takes to build memory a coding agent can actually trust.
This series follows the design of Borg, a PostgreSQL-native memory system that compiles task-specific context from prior coding sessions.
The throughline is one claim: memory for a coding agent is not a search problem, it is a compilation problem. Each article takes one piece of that argument from the core idea down through the architecture, data model, measurements, and research underneath.
Read it in order to follow the whole arc, or jump directly to the part you need: the failure mode, the pipeline, the single-store architecture, temporal facts, benchmarking, or the science behind the design.
Borg is a memory system for AI coding agents: PostgreSQL-native, with task-specific context compilation, temporal facts, and full audit logging.
Frame the agent-memory failure mode
Why stateless coding agents repeat old mistakes, why vector RAG can trade amnesia for stale misinformation, and why memory needs to be compiled instead of searched.
Explain the memory compilation pipeline
How Borg turns a query and prior-session history into a ranked, current, task-shaped context package through an inspectable five-stage memory pipeline.
Defend the single-database architecture
Why a coding-agent memory system can avoid synchronized graph and vector stacks by keeping graph, vector, relational, and audit data inside one transactional Postgres database.
Show how Borg keeps memory current
How temporal facts, lifecycle states, and supersession keep contradicted decisions out of working memory while preserving the audit history.
Explain the benchmark and its limits
How the Borg benchmark compares no memory, vector RAG, and compiled memory across task success, precision, stale facts, irrelevant context, and coverage.
Connect the design to research foundations
How Borg's memory design maps cognitive science, record linkage, bitemporal modeling, and long-context attention research into a practical agent-memory system.