AI Coding Agents Still Forget Everything — So I Built the Memory Layer Underneath Them
AI coding agents are getting very good at editing files, running tests, and opening PRs.
After heavily using tools like Cursor, Claude Code, and GitHub Copilot, I noticed they all share the same core limitation:
They have no persistent understanding of your system.
Ask the same question next week and they:
- re-read the repo from scratch
- re-run expensive LLM calls
- forget prior incidents
- lose architectural context
- and still don’t know what actually happened in production
So instead of building another coding agent, I built the layer underneath them.
Introducing ASIL
ASIL (Engineering Intelligence Infrastructure) is a persistent, temporal, causal knowledge graph for software systems.
It connects:
- code
- commits
- deployments
- incidents
- logs
- metrics
- architecture drift
- AI memory
into one queryable system that any AI agent can access through MCP.
The goal is simple:
Stop making AI agents rediscover the same engineering knowledge over and over again.
The Core Idea
Most coding agents understand:
- the current codebase
ASIL understands:
- how the system evolved
- what changed
- what broke
- why it broke
- and what evidence supports that conclusion
Instead of:
“GPT thinks this caused the outage”
ASIL derives causal chains from observable system state:
- deployment timelines
- incident timestamps
- metric shifts
- runtime dependencies
- postmortems
- service relationships
Every conclusion includes:
- evidence
- confidence scores
- derivation chains
- citations
No black-box “AI intuition.”
What ASIL Can Do
Ask Questions About a Repo
uv run asil ask "How does auth work in this repo?"
ASIL combines:
- graph retrieval
- vector search
- verifier passes
- episodic memory
to return:
- cited answers
- confidence scoring
- cached reasoning for future sessions
Replay Production Incidents
uv run asil replay INC-2026-04-12
ASIL reconstructs:
- deployment timelines
- causal chains
- affected services
- architecture drift
- metric changes
as a dependency-aware replay graph.
Think:
Time-travel debugging for distributed systems.
Detect Architecture Drift
uv run asil drift report
ASIL learns expected dependency boundaries and flags:
- undocumented coupling
- boundary violations
- dependency creep
before the PR merges.
Work With Any Coding Agent
ASIL exposes 13 MCP tools usable from:
- Cursor
- Claude Code
- OpenHands
- Aider
- or custom agents
The agents become clients of the intelligence layer.
The Unexpected Benefit: Massive LLM Cost Reduction
ASIL stores every verified engineering conclusion in persistent memory.
When someone asks a semantically similar question later, ASIL can reuse the prior verified reasoning instead of re-running the full LLM pipeline.
On cache hits, the cost drops close to:
just the embedding lookup
Repeated engineering queries become dramatically cheaper over time — especially across teams.
The Part I Care About Most
ASIL does not let the LLM invent causality.
That rule shapes the entire architecture.
Causal links come from deterministic signals:
- temporal proximity
- lagged correlation
- explicit references
- runtime graph relationships
The LLM consumes evidence.
It does not fabricate it.
That distinction matters once AI systems start participating in production engineering workflows.
Built for Local-First Engineering
Everything runs locally:
- Neo4j
- Qdrant
- Postgres
- Redis
- Prometheus
- Loki
- Grafana
No central server.
No telemetry.
Your graph stays yours.
The only optional network dependency is the reasoning LLM.
Why I’m Building It Open Source
Most AI tooling is racing toward:
“make the agent better at editing code”
I think the more important problem is:
“give agents persistent engineering intelligence”
That means:
- memory
- causality
- runtime awareness
- architecture understanding
- confidence-weighted reasoning
- reproducible evidence
That’s the layer ASIL is trying to build.
Built solo over 6 months with:
Python, FastAPI, Neo4j, Qdrant, Postgres, Tree-sitter, Next.js, Tailwind, ReactFlow, and MCP tooling.
Top comments (0)