hyper-memory skill
✓ VerifiedExtracts and curates evidence-anchored knowledge candidates from local .hyperclaude artifacts into markdown files.
- GitHub stars
- 5
- Files
- 1
- Last commit
- Nov 9, 2026
- License
- MIT
- Allowed tools
- —
What does the hyper-memory skill do?
hyper-memory extracts and curates evidence-anchored knowledge candidates from local .hyperclaude artifacts into markdown files.
Summary generated from the skill's SKILL.md.
About the hyper-memory skill
Extracting durable knowledge from accumulated project artifacts is what hyper-memory enables for developers working with local .hyperclaude files. When an agent detects a significant batch of archived plans, plan-reviews, or research artifacts, it can run this skill to mine those files for evidence-anchored knowledge candidates. The process uses a specific extraction script to scan directories like plans/done, plan-reviews, and research to produce markdown files.
Each generated candidate is stored in .hyperclaude/memory/candidates/ and includes a verbatim evidence section to ensure every claim is traceable to its original source. Developers use this to curate a repository-local memory by moving files from the candidates folder to a promoted folder after adding real repository paths to the anchors list. The tool is designed for orchestration only and does not automatically inject knowledge into future sessions, focusing instead on creating a structured, idempotent collection of verified information.
What an agent can do with it
- 1Mine research for knowledgeScan all accumulated research artifacts to create markdown files containing verifiable claims.
- 2Preview extraction resultsCompute potential knowledge candidates without writing any new files to the disk.
- 3Promote verified informationMove a reviewed candidate into the promoted directory after linking it to canonical repository files.
Written from the skill's SKILL.md. Every use case is checked against a quote from it.
How to install the hyper-memory skill
Each command copies the skill folder into the client's skills directory. Restart the client or start a new session to load it.Install the hyper-memory skill in Plugin
/plugin marketplace add zeikar/hyperclaude /plugin install hyperclaude@hyperclaude
Install the hyper-memory skill in Claude Code
~/.claude/skills/hyper-memory/SKILL.md
$ npx degit zeikar/hyperclaude/skills/hyper-memory ~/.claude/skills/hyper-memoryInstall the hyper-memory skill in Claude Code (project)
.claude/skills/hyper-memory/SKILL.md
$ npx degit zeikar/hyperclaude/skills/hyper-memory .claude/skills/hyper-memoryInstall the hyper-memory skill in Codex
~/.agents/skills/hyper-memory/SKILL.md
$ npx degit zeikar/hyperclaude/skills/hyper-memory ~/.agents/skills/hyper-memoryInstall the hyper-memory skill in Gemini CLI
~/.gemini/skills/hyper-memory/SKILL.md
$ npx degit zeikar/hyperclaude/skills/hyper-memory ~/.gemini/skills/hyper-memoryInstall the hyper-memory skill in Claude.ai
npx degit zeikar/hyperclaude/skills/hyper-memory hyper-memory && zip -r hyper-memory.zip hyper-memory # claude.ai → Customize → Skills → + → Create skill → Upload a skill → hyper-memory.zip
Automated checks
Ran Nov 16, 2026Full report →Passed our automated checks. That is not a security audit: read the scripts before you let an agent run them. How checks work
- Source repository linked and reachablezeikar/hyperclaude/skills/hyper-memory is public and reachablegithub.com/zeikar/hyperclaude/tree/HEAD/skills/hyper-memory ↗Pass
- Open source license (OSI)MITspdx.org/licenses/MIT.html ↗Pass
- Commit or release in the last yearLast commit 2026-11-09github.com/zeikar/hyperclaude/commits ↗Pass
- No download-and-run or credential theft in scriptsNo scripts in the skill folderPass
- No hidden instructions in SKILL.mdNo override phrases, safeguard switches, credential requests or hidden charactersgithub.com/zeikar/hyperclaude/blob/HEAD/skills/hyper-memory/ ↗Pass
- Pre-approved tools are scopedNo pre-approved tools; the agent asks before using onegithub.com/zeikar/hyperclaude/blob/HEAD/skills/hyper-memory/ ↗Pass
- SKILL.md follows the Agent Skills formatValid name and description; says when to use it; 81 linesgithub.com/zeikar/hyperclaude/blob/HEAD/skills/hyper-memory/ ↗Pass
SKILL.md
View on GitHubhyper-memory
Repo-local knowledge extraction. Scans the accumulated .hyperclaude/ corpus and writes one evidence-anchored candidate markdown file per deterministic copy-based span under .hyperclaude/memory/candidates/. v1 is extraction + curation only — auto-injection into future sessions is the v2 north star and is explicitly out of scope here.
When to use
- User typed
/hyperclaude:hyper-memory(with or without an argument). - A batch of work has accumulated in
.hyperclaude/(several archived plans, plan-reviews, research artifacts) and it's worth mining for durable repo-local knowledge.
When to skip
- Only a single small artifact exists since the last extraction — not enough accumulated corpus to be worth mining.
- You want the knowledge injected automatically into a session — that's v2, not implemented.
How it works
-
Run
node "${CLAUDE_PLUGIN_ROOT}/scripts/memory/extract.mjs"via Bash and parse the one-line JSON summary it prints to stdout:{ ok, scanned, candidates, written, skipped, errored, candidatesDir }.The script's CLI accepts exactly two flags — no others exist:
--dry-run— compute candidates and keys but write nothing (writtenis always0).--root <path>— corpus root to scan (default.hyperclaude).
-
It fully enumerates the v1 source allowlist — NOT newest-only:
plans/done/— every archived plan.plan-reviews/— every plan-review artifact whose verdict isShip as-is.research/— every research artifact.
code-reviews/anddocs-reviews/are v1 non-goals and are never scanned. -
It writes one evidence-anchored markdown file per candidate under
.hyperclaude/memory/candidates/, keyed by a compound hash so re-runs are idempotent: a candidate is skipped if its key already exists in EITHER.hyperclaude/memory/candidates/OR.hyperclaude/memory/promoted/— an already-promoted candidate is never resurrected.
Candidate schema
Each candidate file's YAML frontmatter carries exactly these keys, in this order:
plugin-version, type, source-artifact, anchors, mode, slug, git-head, generated, staleness
followed by a ## Claim (a deterministically templated one-liner) and a ## Evidence section holding strictly the verbatim copied span — never a generated or derived line.
CORE POLICY: artifact sentences are never stored as truth on their own — every candidate carries an inline evidence anchor quoted verbatim from source-artifact:, so a claim can always be traced back to the exact text it came from.
Two fields are easy to conflate and must be read as distinct:
source-artifact:— the.hyperclaude/**artifact path the candidate was mined from (evidence provenance; a gitignored artifact, not a canonical repo source).anchors:— a YAML list of live canonical repo source/doc paths the claim is about. The extractor ALWAYS emitsanchors: []— none of the three v1 sources deterministically names a real repo file, and a.hyperclaude/**path is NEVER a validanchors:entry.
Curation
Two locations only — no multi-state machine:
.hyperclaude/memory/candidates/— proposed, unreviewed..hyperclaude/memory/promoted/— human-accepted.
Promote: plain mv .hyperclaude/memory/candidates/<file> .hyperclaude/memory/promoted/<file> — NOT git mv (.hyperclaude/ is gitignored, so git tracks neither side).
Promotion gate: every candidate ships with anchors: []. Before promoting, the curator MUST add at least one real repo source/doc path (a non-.hyperclaude/ file that exists on disk) to the candidate's anchors: list. source-artifact: provenance alone never satisfies this gate — it names a gitignored artifact, not a canonical anchor.
Reject: rm the candidate file.
Idempotency: because promotion is a plain move out of candidates/, and the extractor checks BOTH candidates/ and promoted/ for an existing key, a promoted candidate is never re-created by a later extraction run.
Invocation argument
Invocation argument: $ARGUMENTS
Accepted argument grammar — nothing outside this table:
| Token(s) | Meaning |
|---|---|
--dry-run |
compute candidates/keys, write nothing |
--root <path> |
corpus root to scan (default .hyperclaude) |
The script's CLI parser rejects anything else (unknown flags, --root with a missing/flag-like value) with {"ok":false,"error":...} and a non-zero exit — see scripts/memory/extract.mjs.
Do NOT interpolate the raw $ARGUMENTS string into the Bash command. Parse it into individual tokens, keep only tokens matching the grammar above, and pass each as its own shell-quoted argument (e.g. quote the --root path value) when invoking node "${CLAUDE_PLUGIN_ROOT}/scripts/memory/extract.mjs". Any token outside the grammar means: do not pass it through — the script would reject it anyway.
Files in this skill
1hyper-memory skill questions
How do I install the hyper-memory skill in Claude Code?
Run npx degit zeikar/hyperclaude/skills/hyper-memory ~/.claude/skills/hyper-memory. Claude Code loads skills from ~/.claude/skills (all projects) or .claude/skills (one project) and uses a skill when its description matches the task.
Which clients can use the hyper-memory skill?
Any client that reads the Agent Skills format: Claude Code, Codex, Gemini CLI and claude.ai (upload as a ZIP). Scripts in the skill run only where the client allows code execution.
What files are scanned during the extraction process?
The tool scans archived plans in plans/done/, plan-review artifacts with a 'Ship as-is' verdict in plan-reviews/, and all research artifacts in research/.
How do I promote a knowledge candidate?
You can promote a candidate by moving the file from .hyperclaude/memory/candidates/ to .hyperclaude/memory/promoted/. Before doing so, you must add at least one real repo source or doc path to the anchors list.
Can I run a dry run to see what would be written?
Yes, you can use the --dry-run flag to compute candidates and keys without writing any files.