Drawn directly from the operator's own Wyrm memory, read by their CLI AI assistant, the same way Wyrm is used every working day. Identifying details withheld.
The operator
A forward-deployed operator, running more than a dozen client and product workstreams at once, under constant deadline pressure. One person doing the work of a small team, with an AI assistant as the force multiplier.
The bottleneck was never capability. It was memory. Every new session with the AI started cold.
The challenge: AI amnesia at scale
A model's context window is brilliant but ephemeral. It resets every session and compresses once it fills. Across many parallel projects, that means the thread quietly slips.
- Cold starts. Every session began by re-explaining context from scratch
- Lost decisions. Why a thing was done one way evaporated between sessions
- Repeated mistakes. The same dead ends got tried again, days apart
- Context switching. Thirteen projects, no shared memory across any of them
A real moment
08:00 session start
operator opens the AI. no re-briefing.
wyrm:
picking up where we left off.
63 open threads across 13 projects.
yesterday you were mid way on the migration.
a recorded failure here: do not retry the old auth path, it broke last time.
operator keeps going.The setup: you talk to it, you never manage it
Wyrm is AI-native. There is no dashboard to tend. Your coding AI reads and writes it in plain language, automatically.
Local first
Runs on the operator's machine as an MCP server, backed by SQLite. The memory lives with them.
Plugs into the AI
Claude Code first, and also Copilot, Cursor, Windsurf and Codex. One memory across all of them.
Auto loads
A session-start hook primes Wyrm at the top of every session, before a word of work.
Reflexive
A standing instruction, consult memory first, makes the AI read and write Wyrm on its own.
Working memory plus long term memory
Claude, the working memory
Brilliant, but ephemeral. The context window resets each session and compresses once it fills. On long or parallel work, the thread slips and the assistant loses the plot.
Wyrm, the long term memory
The durable layer beneath it. What the AI learns in a session is written down. Next session it reads it back. Decisions, failures and open threads all survive the reset.
What it saves you
- Lossless rehydration. Start each morning mid stride, not from zero
- Counter-pattern memory. A failed approach is recorded and the AI is blocked from suggesting it again
- Quests. Paused threads resume exactly where they stopped
- Ground truths. The AI never works from a contradicted fact, cascade-aware
- Recall and hybrid search. Ask in plain language, the relevant past work surfaces
The results
77 sessions of continuity. 257 threads resumed. 13 projects in one memory. 450+ distilled know how.
- Context held across 77 sessions and 13 projects, with no cold re-briefing
- 257 paused threads resumed exactly where they stopped
- Recorded failures stopped the same dead ends from being retried
- Mornings started with the full picture loaded in one call
Remembers everything, exposes nothing
Wyrm is local first, and when you sync, the keys are yours. The system is built so the vendor cannot read your memory, by design. For an operator handling sensitive work, that is the difference between a convenience and a liability.
Tech stack
Claude, Claude Code, MCP, SQLite, hybrid search, zero-knowledge sync