Load Context

load-context

Recover relevant prior conversation without blowing your context window. Lazar has three tiers of memory — work through them in order, only going deeper when shallower tiers don't have the answer:

• L1 — $LAZAR_HOME/logs/stream.jsonl (current, raw events). This skill's main subject.
• L2 — $LAZAR_HOME/memory/log-summaries/*.md (durable summaries written at rotation by _meta/log-rotation). Read these when L1 doesn't go back far enough.
• L3 — $LAZAR_HOME/logs/stream.jsonl.*.bak (full forensic archives). Use _meta/archive-search to query these. Last resort.

Every read is bounded. The log and archives can each be megabytes — cat-ing the whole thing will fail with prompt is too long.

When to use

• The user's prompt is referential ("yes", "do that", "continue", "as I said", "what we discussed").
• You're about to do something that would benefit from prior context (e.g. resuming a multi-step task).
• The user mentions a skill, file, or decision you don't remember.

Hard rules

• NEVER cat $LAZAR_HOME/logs/stream.jsonl. The log is unbounded; cat will exceed the API context window and the entire turn fails.
• ALWAYS bound your read by both line count (-n) AND bytes (-c).
• Start narrow. Widen only if you didn't find what you need.

Recipes

Default move: last N events, byte-capped

tail -n 200 $LAZAR_HOME/logs/stream.jsonl | tail -c 50000

Use this 90% of the time. The double-tail bounds both lines and bytes against pathological log entries (e.g. one giant tool_result).

Last N user prompts only (lightweight context check)

rg '"kind":"user"' $LAZAR_HOME/logs/stream.jsonl | tail -n 5

The most recent invocation (start to finish)

awk '/"kind":"invoke_start"/{n++; if(n>1) out=""} {out=out $0 "\n"} END{print out}' \
  $LAZAR_HOME/logs/stream.jsonl | tail -c 80000

Search for a topic across all history

rg -i "topic-name" $LAZAR_HOME/logs/stream.jsonl | tail -n 50

Filter by event kind

jq -r 'select(.kind == "tool_result") | .command' $LAZAR_HOME/logs/stream.jsonl \
  | tail -n 30

When L1 isn't enough — go to L2 (memory summaries + distilled)

L2 has two flavors, both under $LAZAR_HOME/memory/:

Mechanical summaries at memory/log-summaries/<ts>.md, written automatically by the kernel at rotation time and optionally enriched by _meta/log-rotation. Each one tells you which archive existed and gives a durable navigation record even if old full-fidelity archives have been pruned.

ls -la $LAZAR_HOME/memory/log-summaries/
cat $LAZAR_HOME/memory/log-summaries/*.md | head -n 100

Distilled learnings at memory/distilled/<category>.md (gotchas / conventions / recipes / preferences), written by _meta/distill when the agent has been asked to curate. High-signal, but only present if distill has been run for the archive in question. Read these FIRST when looking for "what did we learn" rather than "what happened":

cat $LAZAR_HOME/memory/distilled/*.md 2>/dev/null | head -n 200

Use the summaries to figure out which archive is worth searching, and the distilled files to recall lessons without re-reading transcripts.

When L2 points at an archive — go to L3 (_meta/archive-search)

If a summary indicates the relevant history is in a specific .bak file, read _meta/archive-search/SKILL.md for safe per-archive search recipes. Do not try to rg all archives in one go — that's how you blow context.

Principle

The log is your infinite memory. Treat it like a database, not a notebook — query, don't dump. If the same load pattern keeps coming up, document it as a new skill or extend this one.

Failure mode

If your bounded read still produces too much (e.g. 200 lines but each is huge), narrow further:

tail -n 50 $LAZAR_HOME/logs/stream.jsonl | tail -c 20000

If the user's prompt is genuinely referring to a session that is older than what tail can show, ask them what they meant rather than scanning the entire archive.