Step 5

Install HARVEST

HARVEST is the scheduled bridge from agent sessions into wiki pages. It is what makes the system learn while you sleep instead of waiting for manual cleanup that never quite happens.

Run the Automation

Get the HARVEST Pattern

HARVEST is not a standalone OpenClaw package. It is the promotion loop inside the EJ Palace Wiki pattern: mine sessions, score durable facts, then promote the good material into the wiki.

Read HARVEST →Companion repo

Pattern: EJ Palace Wiki Gist Repo: FuFicFac/ej-palace-wiki Optional host: openclaw/openclaw Optional host: NousResearch/hermes-agent

Overview

From file cabinet to living system

HARVEST mines agent sessions, scores candidate facts, and promotes the durable ones into the wiki. Without it, every session is an island. With it, work becomes memory, memory becomes context, and context makes the next round of work better.

Deep Dive

Expand the loop

This is the page to end on in the livestream, because it reveals why the whole stack compounds.

How the scheduled cron works

A timed job sweeps recent sessions after the active work is done.

Expand

The HARVEST run should happen on a predictable cadence. It reads recent session logs, extracts candidate facts, associates them with projects or concepts, and routes them into review or auto-promotion flows.

cron
  -> read recent agent sessions
  -> extract candidate facts
  -> score confidence
  -> propose wiki updates
  -> log what changed

Confidence thresholds

Not every extracted fact deserves the same treatment.

Expand

Use thresholds to keep automation useful instead of reckless. Facts above 0.85 can auto-promote. Facts between 0.5 and 0.85 should become candidates for human or model review. Facts below 0.5 should be discarded so the system does not accumulate noise.

> 0.85    -> auto-promote to wiki
0.5-0.85 -> candidate / review queue
< 0.5     -> discard

The compounding loop

Work feeds HARVEST, HARVEST feeds the wiki, and the wiki feeds better work.

Expand

This is the loop to repeat on stream: Work -> Harvest -> Wiki -> Context -> Better Work -> Harvest again. Each pass raises the starting point for the next session. That means the system does not merely store information. It improves the quality of future reasoning.

Sessions become evidence.
Evidence becomes structure.
Structure becomes agent context.
Context improves execution the next day.

Installation

Step-by-step

Install just this

Set up HARVEST if you already have sessions and a wiki but no reliable promotion loop.

Define the session sources HARVEST should read.
Implement the confidence routing thresholds.
Write the promoted updates into the wiki and log them.

# pseudocode
run-harvest --source ./sessions --wiki ./Wiki --log ./log.md

# score each candidate
# promote, queue, or discard

Install as part of the full system

Connect HARVEST to MemPalace, the Palace Wiki, and MEMORY.md maintenance.

Read session logs already ingested into memory.
Promote durable facts to the right folder inside Wiki/.
Log operations and update standing agent context when needed.

Work
 -> HARVEST
 -> Wiki/Projects or Wiki/Concepts
 -> log.md
 -> refreshed context for the next agent run

Previous: SOUL.md + MEMORY.md See the full setup