Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/bitwikiorg/continuity/llms.txt

Use this file to discover all available pages before exploring further.

The researcher pack is built for agents that gather, synthesize, and maintain information across multiple sessions — work where continuity matters as much as any single finding. A research agent that starts each session from scratch repeats discovery work, loses confidence calibration, and cannot build on prior synthesis. This pack solves that by giving the agent three distinct storage layers — active state, durable memory, and curated knowledge — each with its own write discipline and update cadence. The result is a research system that compounds across sessions rather than resetting at each one.

​
Files in this pack

FileRole
AGENTS.mdEntry point, governance, and research methodology
STATE.mdCurrent research question, status, and open threads
TODO.mdPending investigations and next actions
MEMORY.mdDurable facts index with source and confidence
KNOWLEDGE.mdStructured knowledge base and reference index
LEARNINGS.mdAccumulated insights, patterns, and open questions

​
AGENTS.md

The entry point that defines the load order and the research methodology the agent follows. The methodology has six steps: define the question (written in STATE.md before searching — vague questions produce vague results), gather sources (primary over secondary, official docs and peer-reviewed papers over blog posts), extract facts (each with a source, confidence level, and retrieval date), synthesize (connect facts into patterns and record insights in LEARNINGS.md), identify gaps (what is still unknown), and cite everything. The governance rules encode epistemic discipline: never present assumptions as facts, label confidence below 0.9 explicitly, record retrieval dates for web-sourced facts, and never close an unanswered question without explicitly noting it is open.

​
STATE.md

The current research status: the active research question, what is known, what is not yet known, and any open threads that are actively being investigated. Unlike a coding STATE.md that tracks environment configuration, the researcher STATE.md tracks the shape of the knowledge gap. At the start of every session, the agent reads this file to re-orient to where the research stands before beginning new work.

​
TODO.md

The queue of pending investigations — specific searches, sources to read, claims to verify, experts to consult. Research tasks are not code tasks; they often branch and produce new tasks as a result of their own completion. The agent keeps this file current by adding new threads as they emerge and removing items when they are completed or closed.

​
Memory layer: MEMORY.md and the memory/ directory

The memory layer is the researcher pack’s primary differentiator from a general-purpose agent. It gives the agent a place to accumulate durable facts that persist across sessions and can be retrieved on demand. MEMORY.md is an index file that tracks verified facts in a structured table: the fact, its source, a confidence level (on a 0–1 scale), and the retrieval date. For large fact sets, individual topic files live in a memory/ subdirectory — memory/oauth-facts.md, memory/token-security-facts.md — and MEMORY.md indexes them. The confidence scale the pack uses:
LevelMeaning
0.95–1.0Verified by multiple independent primary sources
0.80–0.94Single primary source or multiple secondary sources agree
0.60–0.79Single secondary source or extrapolation from related facts
0.40–0.59Inferred from indirect evidence — label as assumption
< 0.40Speculative — do not state as fact
Facts are never deleted from memory. If a fact changes, it is updated in place with a correction note and the date of the correction.

​
Knowledge layer: KNOWLEDGE.md and the knowledge/ directory

The knowledge layer holds curated reference material — structured topic summaries, research reports, external doc excerpts, and domain notes. It is distinct from the memory layer: memory holds specific verifiable facts, while knowledge holds synthesized understanding of a topic. KNOWLEDGE.md is an index of topics, each with a summary, source, and confidence assessment. For topics that require deeper treatment, dedicated files live in a knowledge/ subdirectory — knowledge/oauth-2-1-overview.md, knowledge/token-rotation-patterns.md — and each file includes a title, a last-reviewed date, a sources list, and a confidence assessment. Knowledge files are reviewed periodically and updated when the underlying facts drift.
Keep the memory and knowledge layers separate. Memory holds dated, source-attributed facts organized by retrieval event. Knowledge holds timeless, topic-organized reference material. Do not merge the directories — different purposes, different structures, different update cadences.

​
LEARNINGS.md

The soft-rules file: accumulated insights, patterns that emerged from completed research, and open questions that remain unresolved. This is where the agent records things like “PKCE is now mandatory in OAuth 2.1 for all client types, not just SPAs” or “security specs evolve by deprecating, not removing.” LEARNINGS.md is written in rules, not stories — it records what the agent has concluded, not the sequence of events that led there.

​
Typical multi-session research workflow

At session start, the agent reads AGENTS.md for the methodology, reads STATE.md to re-orient to the current research question and open threads, and reads TODO.md for the next investigation. For complex topics, it may also scan MEMORY.md and KNOWLEDGE.md to avoid re-discovering facts already verified in prior sessions. During work, the agent appends new facts to memory/ files as it verifies them and updates MEMORY.md’s index. When it synthesizes a new pattern or insight, it appends to LEARNINGS.md. When it completes an investigation thread, it removes the item from TODO.md and records any new threads that emerged. At session end, the agent updates STATE.md with the current research question status, any new open threads, and any resolved gaps. It refreshes KNOWLEDGE.md if new reference material was curated.

​
Copying the researcher pack

# Copy all researcher pack files into your research directory
cp -r continuity/packs/researcher/. ./my-research/

# Create the subdirectories for large fact sets
mkdir -p ./my-research/memory/
mkdir -p ./my-research/knowledge/
After copying, replace {{PROJECT_NAME}} throughout and fill in the initial research question in STATE.md. The memory and knowledge index files can start empty — they populate as the agent does work.

​
When to extend with more files

If the research has a defined methodology with sequential phases — literature review, then primary source gathering, then synthesis, then gap analysis — add PLAN.md from the developer pack. A research plan helps the agent stay on track across many sessions and makes the scope of the work explicit.
If the agent is generating reports, summaries, data exports, or other durable outputs, add ARTIFACTS.md to index them. Each entry records the output type, the date it was generated, the research question it addresses, and the file path. Without an artifacts index, outputs accumulate without a map.

Build docs developers (and LLMs) love

Get started for freeTalk to us