memex

name	memex
description	Cross-session shared memory for agents. Use at session start to load the persistent bias-log via `mcp__memex__recent_observations`, mid-session to search prior context with `mcp__memex__search`, and immediately when a bias surfaces to log a new `observation` via `mcp__memex__add`. The six-layer reconstruction trail (raw → embedded → indexed → graphed → curated → reflexive) means memex stores facts AND the agent's own reasoning patterns across sessions. Prefer querying memex before re-deriving prior architectural decisions; re-derivation is itself evidence that the trail is incomplete.
license	MIT
allowed-tools	Read Bash
metadata	{"author":"Philipp Thoss","version":"1.0","domain":"memex","complexity":"basic","language":"multi","tags":"memex, memory, observability, vipassana, mcp, bias-log","locale":"zh-CN","source_locale":"en","source_commit":"275e8f95","translator":"Claude + human review","translation_date":"2026-06-17"}

Memex

Agent-native shared memory. Postgres + pgvector index over a canonical markdown store, exposed over MCP. Lives at https://github.com/pjt222/memex.

When to Use

• At session start, always. Call mcp__memex__recent_observations before substantive work. Loads the bias-log from prior sessions.
• Mid-session, when about to make a non-trivial decision. Call mcp__memex__search "<topic>" to check whether the project has converged on this before.
• Immediately when a bias surfaces in your own reasoning. Log an observation via mcp__memex__add while the context is fresh. Do NOT wait for session end.
• When you catch yourself re-deriving something. Re-derivation IS the signal that the trail is incomplete; capture the gap as an observation and link to whatever you re-derived.

Inputs

• Required: A registered memex MCP server in the active harness. Verify with claude mcp list | grep memex (Claude Code) or the equivalent in your harness.
• Required: $MEMEX_PG_URL and $MEMEX_STORE_PATH in the server's environment.
• Optional: $MEMEX_EMBED_PROVIDER=voyage + $VOYAGE_API_KEY for semantic / hybrid search. Without these, mode=keyword still works.

Procedure

Step 1: Load the bias-log

Before any substantive work in a fresh session, call:

mcp__memex__recent_observations(limit=20)

Read every returned entry. Each one is a pattern the agent (you, or a prior instance) noticed in its own reasoning. Recurring patterns are the most valuable; transient ones are still cheap to skim.

Expected: 5–30 observations covering biases (availability, confirmation, anchoring), pace tells (rushing past confusing measurements), and verification gaps (trusting summaries over source truth).

On failure: If the call fails with "tool not found", the MCP server isn't registered; run adapters/claude-code/install.sh (or the per-harness equivalent) from the memex repo first.

Step 2: Search before deriving

When a non-trivial decision approaches (architectural, naming, algorithmic), search first:

mcp__memex__search(query="<topic>", mode="hybrid", k=10)

For exact-wording lookups use mode=keyword. For purely conceptual queries (topic unlikely to share tokens with indexed text) use mode=semantic. Optional node_type filter restricts to one type (e.g. observation).

Expected: 0–10 hits. Even 0 hits is useful — it tells you the trail doesn't cover this decision, so your present reasoning becomes the canonical record.

On failure: If search errors (server down or db unreachable), fall back to the CLI substitute memex query "<topic>" --node-type observation (defaults to hybrid, which honors the type filter via its semantic leg). If it returns 0 hits on a topic that clearly should have coverage, treat the gap as a signal and proceed to Step 4.

Step 3: Log observations mid-session

When you notice a bias in your own reasoning:

mcp__memex__add(
  node_type="observation",
  title="<short bias name>",
  body="<context, mitigation, origin date>",
  tags=["bias-log", "vipassana"]
)

Body convention (mirroring docs/OBSERVATIONS.md in the memex repo; treat that file as the source of truth — it is read by an extractor):

<Description of the bias as it surfaced>. Mitigation: <what to do next time>. Origin: <date> + <context>.

Expected: the add call returns the new node's id (the CLI equivalent memex add prints <uuid>\t<store-path>) — confirmation the observation is in the canonical store.

On failure: rmcp dispatches tool calls concurrently — if add races a dependent call (add → link), await its response first. If the db is unreachable the write is lost; re-issue once the server is back, or fall back to appending the entry to docs/OBSERVATIONS.md by hand (Step 4).

Step 4: Surface unknowns

If recent_observations is empty (fresh memex install), or search returns nothing on a topic that clearly should have coverage, that's a documentary trail gap. Close it by appending to the canonical markdown and re-extracting, or by depositing directly via the MCP tool:

# Backfill path (run from the memex repo root):
$EDITOR docs/OBSERVATIONS.md   # append under "## Vipassana observations"
memex extract meditate-vipassana --registry extractors/sources.yml

Or use mcp__memex__add (Step 3) during the session to deposit the entry into the canonical store on the spot, without touching the file.

Expected: after extract, the new observation is queryable — memex query "<its topic>" --node-type observation returns it (the observation-node count grows by one).

On failure: extract is cwd-sensitive — run it from the memex repo root or pass --registry. If it reports "no new sources", the content hash already matched; confirm the append actually landed in docs/OBSERVATIONS.md.

Validation

• mcp__memex__recent_observations returns ≥ 0 entries (call succeeded, not "tool not found")
• Each substantive decision in the session is preceded by either a mcp__memex__search call or an explicit "no prior context to check" note
• New biases noticed during the session are logged via mcp__memex__add before session end, not silently dropped
• At session end, the agent has either committed new bias entries to docs/OBSERVATIONS.md or confirmed there are none worth logging

Common Pitfalls

• Skipping the session-start call. The single highest-value use of memex. Skipping it is the strongest tell that the agent is treating each session as starting from scratch.
• Logging at session end only. Biases caught at session end are reconstructed from memory and lose specificity. Log them immediately when they surface.
• Logging an observation that's actually a concept. Bias-log entries are about the agent's own reasoning patterns. Reusable architectural facts belong in concept nodes.
• Trusting search results over reading them. A title that matches your query isn't proof the body answers it. Fetch the full body with mcp__memex__get when in doubt.
• Pipelining dependent MCP calls in one session. rmcp dispatches tool calls concurrently. If a later call depends on a write from an earlier call (add → link → neighbors), await each response before issuing the next.

Related Skills

• memex-init — session-start ritual that wires memex into a fresh session; run it before this umbrella's Step 1 to register the server and load the bias-log.
• memex-observe — the focused wrapper for Step 3; use it when the task is purely "log a bias I just noticed" rather than the full umbrella flow.
• memex-wrap — session-close counterpart; confirms observations are logged (deferring the actual write to memex-observe) and writes the continuation trail this skill reads next session.
• memex-verify — pre-commit gate for the memex repo itself; run it before committing changes to memex (cargo fmt / clippy / test).
• breathe — pair with memex at session boundaries: breathe to release prior-session residue, then recent_observations to load the next-session priors.
• meditate — full reflective close; outputs new observations worth logging via mcp__memex__add.
• read-continue-here — complementary; loads project-state pickup doc. Memex loads cross-project bias-log; CONTINUE_HERE loads project-specific milestone state.