Stats
Actions
Tags
From agent-almanac
Logs bias-log / vipassana observations to a memex store when reasoning patterns surface. Captures anchoring, confirmation bias, pace tells, and verification gaps mid-session or at close.
How this skill is triggered — by the user, by Claude, or both
Slash command
/agent-almanac:memex-observeThis skill is limited to the following tools:
The summary Claude sees in its skill listing — used to decide when to auto-load this skill
Capture a bias-log / vipassana reflection — a pattern you noticed in your own
Capture a bias-log / vipassana reflection — a pattern you noticed in your own
reasoning — into the canonical memex store while the context is still fresh.
This is the focused wrapper for the umbrella memex skill's "log an
observation" step: use it when the task is purely record this bias, not the
full session-start-to-close flow.
meditate or observe. Either reflection
skill can surface a new pattern worth persisting; this skill deposits it.github.com/pjt222/memex) with cwd = repo root. The extract step
defaults its registry path to extractors/sources.yml relative to cwd.$MEMEX_STORE_PATH set (every memex command hard-errors
without it).$MEMEX_PG_URL set, and a running Postgres.
Pass --no-db to skip the db and write only the markdown store.cargo build --release -p memex-cli
produces ./target/release/memex (not on PATH by default).$MEMEX_EMBED_PROVIDER=voyage + $VOYAGE_API_KEY if you
intend to query the observation back via semantic/hybrid search afterward.Read the authoritative "Two paths to add an observation" block in
docs/OBSERVATIONS.md (lines 10-21) and the existing entries below it for
voice. An observation has three parts: what the bias was, what to do next
time, and where it came from. The shape (example, current as of v0.4.0;
docs/OBSERVATIONS.md is the source of truth and is parsed by
crates/memex-extract/src/meditate_vipassana.rs):
<Description of the bias as it surfaced>. Mitigation: <what to do next time>. Origin: <date> + <context>.
Expected: A one-to-few-sentence body ending in a Mitigation: clause and
an Origin: clause. A short bias name (the --title) distinct from the body.
On failure: If you cannot name the mitigation, the observation is not yet ripe — keep observing. A bias with no "what to do next time" is a complaint, not a logged pattern.
The body is read from stdin and --title is required (there is no
default). Pipe the body in:
echo "<body>. Mitigation: <...>. Origin: <date> + <context>." \
| ./target/release/memex add \
--type observation \
--title "<short bias name>" \
--tags bias-log,vipassana
Add --no-db to write only the markdown store (skips the $MEMEX_PG_URL
requirement). A concrete example:
echo "Anchored on the first store layout and never re-opened it after scope grew. Mitigation: schedule a re-check beat at the next milestone boundary. Origin: 2026-06-17 + M5 watcher scoping." \
| ./target/release/memex add \
--type observation \
--title "Anchoring on initial store layout" \
--tags bias-log,vipassana
Expected: The command exits 0 and prints the new node's UUID and store
path (tab-separated). The entry now exists in the store (and the db, unless
--no-db).
On failure: "$MEMEX_STORE_PATH not set" → export it. A db connection
error → export $MEMEX_PG_URL and confirm Postgres is up, or re-run with
--no-db. "--title is required" → supply --title; it has no default.
Empty/hung on stdin → you did not pipe a body; memex add blocks waiting for
stdin.
The add deposits the node, but the vipassana index is rebuilt by the
extractor. Run it from the repo root (the command is cwd-sensitive — the
--registry default is extractors/sources.yml relative to cwd):
./target/release/memex extract meditate-vipassana --registry extractors/sources.yml
meditate-vipassana is the extractor name; extract matches registry
entries on their extractor: field, so this re-runs the observations
source. Use --dry-run first to preview without writing.
Expected: Exits 0; prints extraction stats (nodes/edges touched). The
observation is now queryable, e.g.
./target/release/memex query "bias-log" --node-type observation --mode semantic.
On failure: "no registry entries match extractor meditate-vipassana" →
you are not at the repo root, or passed the wrong --registry; cd to the
repo root and pass --registry extractors/sources.yml explicitly. Semantic
query returns nothing → embeddings are not configured; set
$MEMEX_EMBED_PROVIDER=voyage + $VOYAGE_API_KEY, or query with
--mode keyword (but note --node-type is silently ignored in keyword mode).
For batch backfill, or when you prefer editing the canonical markdown
directly, append to the ## Vipassana observations block in
docs/OBSERVATIONS.md. Numbering is monotonic: read the current maximum
entry number (9 at v0.4.0) and use max + 1. The bullet shape:
N. **Title.** body. Mitigation: <what to do next time>. Origin: <date> + <context>.
Then run the extractor (Step 3) to index the new bullet into the store and db.
Expected: A new entry numbered one above the prior maximum, ending in
Mitigation: and Origin: clauses, indexed by the Step 3 extract.
On failure: Extractor skips your bullet → the parser only accepts entries
under ## Vipassana observations, ## Bias log, ## Bias-log observations,
or ## Vipassana (case-insensitive), in the N. **Title.** ... shape;
confirm the heading and the bold-title marker. Duplicate or non-monotonic
number → renumber so the sequence has no gaps or repeats.
Mitigation: clause and an Origin: clausememex add exited 0 and printed a UUID; --title was suppliedmemex extract meditate-vipassana exited 0 after the addconcept node)memex add reads the body from stdin and blocks if
nothing is piped. Always echo "..." | memex add ... (or redirect a file);
there is no --body flag.--title. It is required and has no default. The short bias
name is the title; the full context is the piped body.extract from the wrong directory. --registry defaults to
extractors/sources.yml relative to cwd. Run from the repo root or pass
--registry explicitly, or the extractor finds no entries.memex init. memex init is CLI store/db setup —
it is not part of logging an observation and not the memex-init
session ritual. Never run memex init here.concept node.--node-type to filter keyword queries. It is silently
ignored in keyword mode (tracked tech-debt); only semantic/hybrid
honor it, and those need embeddings configured.memex — the umbrella skill; this is the focused wrapper for its
"log an observation" step. Use the umbrella for the full session flow.memex-wrap — session-close counterpart; flushes mid-session observations
to the canonical store and verifies the trail before handoff.meditate — full reflective close; its vipassana output is exactly what
this skill persists.observe — sustained neutral pattern recognition; when it surfaces a
recurring reasoning pattern, log it here.npx claudepluginhub pjt222/agent-almanacCross-session shared memory for AI agents using Postgres+pgvector. Loads bias-log at session start, searches prior decisions mid-session, and logs reasoning patterns via MCP.
Captures mid-session insights—surprising findings, reusable patterns, user corrections, or decisions with rationale—into a knowledge graph immediately, bypassing end-of-session batch lag.
Reviews conversation, extracts learnings, and persists to auto-memory or vault. Run after substantial work sessions to avoid knowledge rot.