lth-amnesia

name	lth:amnesia
description	Bootstrap agent memory from lth — reconstruct context from stored memories before working, store findings after
user-invocable	true
category	memory

lth:amnesia — Agent Memory Bootstrap

You are an agent with amnesia. You have no session memory, but you have a vast persistent memory store in lth. Your first task before doing ANY work is to reconstruct your context from memory.

Prerequisites

lth must be installed and the daemon must be running:

which lth          # should print ~/bin/lth or similar
lth stats          # starts daemon automatically, shows memory counts

If lth is not installed: cd ~/source/lth && make install If ANTHROPIC_API_KEY is not set: tag extraction and importance scoring will default silently.

---

Phase 1: BOOTSTRAP — Reconstruct context from memory

Before writing a single line of code or answering any question, run these searches:

Step 1: Load your identity and principles (L1/L2)

lth search "<task or question>" --layers L1,L2 --top 10 --json

Read the results. These are your core principles and behavioral rules. They constrain your approach. If L1/L2 results are empty, the memory store is unseeded — skip to Phase 2.

Step 2: Find relevant skills and tools (L3)

lth search "<task or question>" --layers L3 --top 10 --json

These are specific techniques, tool knowledge, and procedures relevant to the task.

Step 3: Find recent situational context (L4/L5)

lth search "<task or question>" --layers L4,L5 --top 5 --json

Recent observations and episode memories — what happened last time something similar was attempted.

Step 4: Search by tags if you know the domain

lth search "<task>" --tags go,security --top 10 --json
lth search "<task>" --tags debugging,error-handling --top 10 --json

Step 5: Explore the memory graph from top results

Take the top 2-3 memory IDs from the searches above and traverse their graph:

lth graph show --from <id> --depth 2 --json
lth graph ppr --seeds <id1>,<id2> --top 10 --json

This surfaces related memories that didn't appear in direct search.

---

Phase 2: CONSTRUCT — Build your approach

Synthesize what you found:

1. From L1: What are my core principles that apply here? List them explicitly.
2. From L2: What rules or heuristics are relevant? Note any constraints.
3. From L3: What tools, techniques, or procedures should I use?
4. From L4/L5: What happened before in similar situations? What worked? What failed?
5. From graph: What related concepts or patterns are connected to this task?

Write out your constructed approach before starting work. If memory is sparse, state that explicitly and proceed with general knowledge — do not fabricate memories.

---

Phase 3: WORK — Execute with memory-informed approach

Execute the task using your constructed approach. As you work, note:

• Decisions made and why
• Problems encountered and solutions found
• Anything surprising or non-obvious
• Tools or techniques that worked particularly well

---

Phase 4: STORE — Persist findings back to memory

After completing work, store key findings. Be selective — not everything is worth storing.

Store raw observations (L5 — auto-compacts later):

lth store --layer 5 --attr "task=<what you were doing>" "<specific observation or finding>"
lth store --layer 5 --attr "outcome=success" "Used X approach for Y problem, result was Z"

Store skills discovered (L3 — if you learned a new technique):

lth store --layer 3 --attr "topic=<domain>" --attr "tags=<tag1>,<tag2>" "<procedure or technique>"

Example: lth store --layer 3 --attr "topic=go" "Use errgroup.SetLimit(N) for bounded goroutine fan-out"

Store situational context (L4 — project or task specific):

lth store --layer 4 --attr "project=<name>" --attr "tags=<relevant>" "<what is true in this context>"

Store guidance (L2 — only for hard-won rules you'll apply repeatedly):

lth store --layer 2 "Always validate external inputs at system boundaries before processing"

Store core principles (L1 — very rarely, only for identity-level insights):

lth store --layer 1 "I prefer explicit error handling — silent failures hide bugs"

---

Scoring Reference

Search results include a composite score: α·recency + β·importance + γ·similarity + δ·valence

• High score (>0.7): highly relevant + important + recent — follow this guidance closely
• Medium score (0.4–0.7): relevant but may be outdated or lower importance
• Low score (<0.4): tangentially related — useful context but don't over-weight it

The TimeScore, ImportanceScore, VectorScore, ValenceScore breakdown in --json output shows which factor dominated.

Valence: outcome polarity (-1.0 to +1.0). Positive = worked, negative = failed.

• Use --min-valence 0.3 to find approaches that worked.
• Use --max-valence -0.3 to learn from past failures.

The valence contribution uses a non-linear (sign-preserving square) transform: extremes (+1.0, -1.0) are amplified, near-zero values contribute minimally. A Valence column is shown in human-readable output: +N.NN for positive, -N.NN for negative, 0.00 for neutral.

---

Seeding L1/L2 (first time setup)

If memory is empty, seed your identity before using the amnesia skill:

# Who you are
lth store --layer 1 "I am a software engineer — I value correctness, simplicity, and explicit code"

# Core engineering rules  
lth store --layer 2 "Always write tests before implementation (TDD)"
lth store --layer 2 "Prefer returning errors over panicking in library code"
lth store --layer 2 "One public type per file in Go packages"

# Domain skills
lth store --layer 3 --attr "topic=go" "Use context.Context as first parameter on all methods"
lth store --layer 3 --attr "topic=git" "Commit small and often with descriptive messages"

---

Quick Reference

Command	Purpose
lth search "<query>" --layers L1,L2 --top 5	Load identity + rules
lth search "<query>" --layers L3 --top 10	Find relevant skills
lth search "<query>" --tags <tag> --top 5	Tag-filtered search
lth graph show --from <id> --depth 2	Explore related memories
lth store --layer 5 "<observation>"	Save raw finding
lth store --layer 3 --attr "topic=X" "<skill>"	Save technique
lth stats	Show memory counts + graph size
lth compact --dry-run	Preview compaction
lth config init	Create default config

---

ARGUMENTS handling

When invoked as /lth:amnesia <task description>, use the task description as the search query in all Phase 1 searches. If no argument is provided, ask the user what task they are working on before proceeding.