# Research What we looked at while designing Tank and Duck (October 2026), what we took and what we left. Sources were read to varying depth. Nothing listed here was installed or tested. ## Taken | Source | What we took | | --------------------------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | Amp: Puck, oracle, find/read threads (ampcode.com/news; Thorsten Ball's October 2026 talk) | A home-base agent you talk to. A read-only second-opinion agent on another model, used at checkpoints rather than every turn. Past threads searched on demand instead of summarized up front. | | Claude Code docs: `/recap`, prompt caching, agent view, cross-session messaging, auto memory | A recap that doesn't break the prompt cache. Agent view's grouping (needs input, working, completed), peek and attach. Direct messages between sessions. | | Claude Code internals (leaked-source mirrors and bundle analyses, unofficial) | Side calls that share the parent's prompt cache. Background summaries are not free: issue #86879 measured them at 22.7% of requests in one session. | | Claude Code `/insights` | Record fields: goal, outcome, summary, friction, instructions you keep repeating. | | Codex memories (`codex-rs/memories`) | Extract only when a future agent would act better; redact secrets; consolidate into a git-tracked folder reviewed as a diff. | | Karpathy's LLM Wiki gist; willcb on X | Raw sources stay untouched. A filesystem plus a map beats clever retrieval. | | pi-hermes-memory | Rules have to be loaded every time; facts can be looked up. Rules go in AGENTS.md. | | Herdr | Status model: working, blocked, done but unseen, idle. | | Prime Agent (repo; Seth Karten interview, October 2026) | Direct addressed messages instead of a shared board. Helpers that go idle and resume with their history. Generator and reviewer on different models. Memory edits as small, reviewable, reversible changes (`/refine`). | | pinet | A per-pane state publisher as the base layer. Named agents. Links (PRs, previews) saved per session. | | Observer pattern: Claude Code observer analyses (unofficial), pi-supervisor, the-verifier-agent, LivePlan paper | Reviewer silent by default. Cheap triggers before LLM calls. | | tmustier/vibes | Papercut notes captured in the moment and reviewed later. | | Harness research session (October 26) | A mechanism's value depends on the model and task, and mechanisms don't add up. Giving agents context as files carries most of the RLM gain. | | agent-stuff `control.ts` and `/btw`; pi-subagent | Messaging over per-session sockets already exists. A side session in an overlay that can inject a summary into the main session. Send and follow-up to child sessions. | | pi-mem (jo-inc) | Topic notes read on demand and a git commit per write (not its scratchpad or daily log). Its "last 24h" widget groups recent sessions by topic with an LLM call, which is prior art for the tower's grouping. | ## Left | Source | Why not | | ------------------------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | magic-context, claude-mem, mem0, graphiti, Letta | Agent-written memory, embeddings, replacing compaction. Not our problem at ~50 sessions a week. | | Herdr as a runtime | Means living inside a multiplexer and dropping Ghostty tabs and session restore. (We later moved to Rex, which has native tabs and keeps sessions across app restarts.) | | pinet's broker | Every message goes through one broker; we want direct messages between named sessions. | | Prime Agent's REPL and deep recursion | Built for long autonomous runs, not for your attention across sessions. | | Wiki or dreaming layer | Wait until searching records proves insufficient. | | lat.md | A per-codebase knowledge graph; a different problem. | | Meta-Harness paper | Needs a benchmark score. Kept only the point that raw traces beat compressed feedback. | | Agentic engineering taxonomy (pasted LLM output) | Kept the test for additions (which pain it removes, what it costs, what shows it works) and the handoff fields for records. | | pi-mem as a package | It appends its notes, the last two days of logs and catch-up notes as a trailing user message on every model request, so they are re-sent uncached each time. The system prompt is left alone. We copy the layout, not the package. | | Benchmarks for memory (LongMemEval and similar) | They score recalling facts from long chats. Nothing we found measures notes a meta agent keeps for one person, so the choice rests on control, reviewability and cache stability. |