Skip to content

project-memory-brief.md

Latest commit

 

History

History
86 lines (63 loc) · 2.68 KB

project-memory-brief.md

File metadata and controls

86 lines (63 loc) · 2.68 KB
Raw
project project-memory-demo
brief_id 2026-05-17-PMD
scope_locked true
mode gated
authorized_costs
filesystem_read
filesystem_write
inherit_memory true
budgets
max_tokens max_wall_clock_minutes
30000
10

Project memory demo (v0.6)

Objective

Demonstrate the v0.6 persistent project memory loop end to end:

  1. The agent reads any existing memory at session start via inherit_memory: true.
  2. The agent makes a decision during execution and writes it to decisions via write_project_memory.
  3. The agent adds a constraint that future briefs should respect.
  4. The next brief against the same project: slug sees the entry without operator intervention.

Run this twice. The second run starts with non-empty memory; the agent should notice and reference it.

Deliverables

A single completion report that explicitly:

  • Lists what was in memory at session start (zero on the first run, non-zero on the second).
  • Records one new decisions entry summarising the choice between two equally valid options the agent picks during the run.
  • Records one new constraints entry that captures a rule future briefs should respect.

Both writes use the native write_project_memory tool. No manual file editing.

Constraints

  • Cap on inherit_memory injection is 20 entries per category. Do not assume the full history is visible on a long-running project.
  • Memory is per-project. Cross-project context is out of scope; do not attempt to read another project's memory.
  • Decisions are visible only to planner / reviewer / critic in multi-agent runs. This brief is single-agent so the full view is in play.

Tools required

  • read_project_memory (default-allowed via filesystem_read).
  • write_project_memory (requires filesystem_write, which this brief authorizes).

Notes / unknowns

Run from this repo's root:

export ANTHROPIC_API_KEY=...
npx @pythonluvr/openwar run examples/project-memory-brief.md --adapter anthropic

Inspect the resulting memory between runs:

npx @pythonluvr/openwar memory list project-memory-demo
npx @pythonluvr/openwar memory show project-memory-demo <entry_id>

Reset for a clean second-run demo:

npx @pythonluvr/openwar memory remove project-memory-demo <entry_id>
# or just rm -rf ~/.openwar/projects/project-memory-demo

Multi-agent variant (executor's view omits decisions):

roles:
  planner:
    adapter: anthropic
    model: claude-haiku-4-5
  executor:
    adapter: anthropic
  reviewer:
    adapter: anthropic
    model: claude-haiku-4-5

The reviewer can still raise prior-decision concerns because reviewer has the full view. The executor solves the current sub-task without bias from past decisions.