How I stopped my AI coding agent from being amnesiac

ivncmp — Wed, 03 Jun 2026 12:10:23 +0000

How I built an engineering toolkit for AI coding agents

I spend most of my day working with AI coding agents — Claude Code, Gemini CLI, Codex, OpenCode. They're incredibly powerful, but they all share the same fundamental problems:

They forget everything. Every session starts from zero. Yesterday's decisions? Gone.
They don't know your project. File paths get hallucinated. Architecture gets ignored.
You can't see what they cost. Tokens, tools, errors — all a black box.
Switching providers is painful. Your setup is tied to one vendor.
Team knowledge stays siloed. Each developer's agent is an island.

After months of duct-taping workarounds, I decided to build the infrastructure layer these agents are missing. The result is dtoolkit — an open-source engineering toolkit that gives AI coding agents memory, context, project management, and observability.

The architecture

dtoolkit follows one principle: one layer, one job. Each package does exactly one thing and does it well.

Your coding agent (Claude Code, Gemini CLI, Codex, OpenCode)
        │
        ▼
┌─────────────┬──────────────┐
│  dcontext   │    dproxy    │  ← Hooks & transport
│  (inject)   │  (route)     │
└──────┬──────┴──────┬───────┘
       │             │
┌──────▼───┐  ┌──────▼───┐  ┌────────┐
│  dbrain  │  │  dwork   │  │  dops  │  ← Data & services
│ (memory) │  │(projects)│  │ (obs)  │
└──────────┘  └──────────┘  └────────┘

No package depends on another to function. You can use dbrain alone for memory, dwork alone for project management, or all of them together. SQLite everywhere — no Postgres, no Redis, no Docker.

Getting started (5 minutes)

1. Install

npm i -g @dtoolkit/dbrain @dtoolkit/dcontext @dtoolkit/dwork @dtoolkit/dops

2. Initialize and start the memory server

dbrain init && dbrain start

This creates a personal brain — a SQLite database with FTS5 full-text search, served as both a REST API and an MCP server on port 7878. A React dashboard runs on 7879.

3. Connect to your AI coding CLI

dbrain connect claude

This configures Claude Code's hooks so that:

At session start: your identity, project facts, and brain data are injected into the context
Before compaction: transcripts are saved so nothing is lost

One command. Zero config files to edit manually.

4. Start the project manager

dwork init && dwork start

dwork turns Markdown files into a project management system. Your BACKLOG.md is the source of truth — SQLite + FTS5 is just an index. It includes a kanban dashboard, 21 MCP tools (including code graph analysis), and full-text search across all projects.

5. Open your agent

claude  # or gemini, codex, opencode

Your agent now has memory. It knows who you are, what project you're in, and what you decided last week.

The five products

dbrain — Your distributed mind

The core of the system. A persistent memory server that stores entities, facts, and conversations with a tiered memory system (hot, warm, cold).

Key features:

FTS5 search — semantic search across all stored knowledge in milliseconds
MCP server — works with Claude Code, Cursor, Windsurf, and any MCP client
Federation — personal and shared team brains. recall auto-federates across connections
React dashboard — browse entities, search memories, manage your brain visually

import { DBrainClient } from '@dtoolkit/sdk';

const brain = new DBrainClient('http://localhost:7878');

// Remember something
await brain.remember('user:ivan', 'Prefers TypeScript over JavaScript');

// Recall later
const results = await brain.recall('typescript preferences');

dcontext — Automatic context injection

The glue between your brain and your coding agent. Hooks into the CLI's lifecycle events:

Session start: injects identity, project facts, active tasks
Pre-compaction: saves the transcript before the context window shrinks
Multi-CLI: works with Claude Code, Gemini CLI, Codex CLI, OpenCode

dproxy — Multi-provider transport

One CLI and REST API that routes to any provider:

dproxy ask "explain this function" --provider gemini
dproxy ask "explain this function" --provider claude

Includes a context pipeline that automatically assembles memory, workspace info, and templates into every prompt. SSE streaming, Bearer token auth, configurable port.

dwork — Markdown-driven project management

Your BACKLOG.md looks like this:

## Todo

- [P0] Build authentication module (id:abc123, est:4h)
- [P1] Add dark mode to dashboard (id:def456, est:2h)

dwork parses it, indexes it in SQLite + FTS5, and serves it via REST + MCP. The dashboard gives you a kanban board with drag-and-drop. The code graph feature maps your entire codebase as a knowledge graph — search for functions, trace call chains, analyze impact.

dops — Agent observability

Finally know what your agents cost:

Token tracking per session (input, output, cache)
Model-aware cost estimation across all providers
Tool call analytics and success rates
Multi-provider transcript ingestion
Dashboard with charts and timeseries

The SDK

Everything is accessible through typed clients:

npm install @dtoolkit/sdk

import { DBrainClient, DWorkClient, DOpsClient, DProxyClient } from '@dtoolkit/sdk';

Four clients, consistent API, full TypeScript types.

Design philosophy

Six principles guide every decision:

One layer, one responsibility — memory is memory, projects are projects
CLI-first, dashboard as bonus — everything works from the terminal
SQLite everywhere — zero external dependencies
MCP + REST on the same port — one server, two protocols
Markdown as source of truth — the database is just an index
No vendor lock-in — four adapters, one interface

What's next

dtoolkit has 20 packages total. 12 are available now on npm. 8 more are on the roadmap:

dcouncil — multi-agent debate for architecture decisions
dguard — pre-commit validation for LLM output
dpolicy — policy-as-code for the team harness
droute — model router with cost tracking
dpair — real-time pair programming with shared agents
dreplay — privacy-aware session browser
dstream — daily digest of what agents learned
dforge — marketplace for skills and hooks

Try it

npm i -g @dtoolkit/dbrain @dtoolkit/dcontext
dbrain init && dbrain start
dbrain connect claude
claude

Five commands. Your agent now has a brain.

All MIT licensed. v1.0.0 just shipped. Feedback and contributions welcome.

DEV Community: ivncmp