codex-agent-mem

v1.0.0 verification results

These are reproducible, sanitized results generated from synthetic fixtures.

Execution context:

• Runtime: Codex Desktop in a Codex environment
• Model: GPT-5.4
• Reasoning effort: xhigh

Scope: this validates Codex Desktop, not a ChatGPT web/app connector. Claude web / claude.ai support is also not inferred from these results.

Snapshot

Scenario	Source tokens	Pack tokens	Saved	not_modified	Tools	Lazy init	Read-only
Small project continuity	1,841	216	88.27%	true	4	false->true	true
Medium agent workflow	4,855	233	95.20%	true	4	false->true	true
Large repeated audit	9,731	232	97.62%	true	4	false->true	true
Sub-agent handoff example	6,523	239	96.34%	true	4	false->true	true

Across these fixtures: ~22,950 source tokens -> ~920 pack tokens. Approximate repeated-context reduction: 95.99% (~22,030 tokens not resent).

Token savings by scenario

Small project continuity

A short project where the agent should remember objective, constraints and open work without repeating the whole discussion.

• Source context: ~1,841 tokens
• Memory pack: ~216 tokens
• Tokens not resent: ~1,625
• Estimated savings: 88.27%
• Pack hash: 9c219f1ea5993da6389c683a029e57c40aed129186e9fdfbe9de8058845967d7

source [############################] 100.00% pack [###…………………….] 11.73% saved [#########################…] 88.27%

Medium agent workflow

A realistic multi-step implementation with repeated decisions, pending work and DoD requirements.

• Source context: ~4,855 tokens
• Memory pack: ~233 tokens
• Tokens not resent: ~4,622
• Estimated savings: 95.20%
• Pack hash: 762cd3a6bb36f47b76bf12f629b55e2c9d4c18b7e59a3a94083af580c012775b

source [############################] 100.00% pack [#………………………] 4.80% saved [###########################.] 95.20%

Large repeated audit

A long audit where the same constraints and decisions would normally be re-sent many times.

• Source context: ~9,731 tokens
• Memory pack: ~232 tokens
• Tokens not resent: ~9,499
• Estimated savings: 97.62%
• Pack hash: 2a4983655139737e718c1a4315e1797a805eeb5294eba9e8dd5cb8472bcbad50

source [############################] 100.00% pack [#………………………] 2.38% saved [###########################.] 97.62%

Sub-agent handoff example

A project where an explorer sub-agent audits context and a worker sub-agent implements a bounded change.

• Source context: ~6,523 tokens
• Memory pack: ~239 tokens
• Tokens not resent: ~6,284
• Estimated savings: 96.34%
• Pack hash: 55cfc4345a1ccafeb426c30e451a13fa8da84e928a501b165d8e82d5be7c4e46

source [############################] 100.00% pack [#………………………] 3.66% saved [###########################.] 96.34%

Sub-agent example:

• explorer: Audit the memory surface and identify which facts must never be dropped.
• worker: Implement the bounded runtime-efficiency patch without touching documentation.

Repeated context avoided

known_pack_hash lets the agent ask whether a pack changed before re-sending it.

Scenario	Result
Small project continuity	not_modified=true
Medium agent workflow	not_modified=true
Large repeated audit	not_modified=true
Sub-agent handoff example	not_modified=true

Runtime safety

Metric	Result
Minimal profile tools	mem_open_work, mem_completion_check, mem_context_pack, mem_health_runtime
Tool count in minimal profile	4
Lazy initialization before DB-backed tool	false
Lazy initialization after context pack	true
Mutating tool tested in read-only mode	mem_snapshot_create
Mutating tool blocked	true

Response diet

Text shown to the model can be kept compact while the structured payload remains available to MCP clients.

Scenario	Compact text chars	Balanced text chars	Verbose text chars
Small project continuity	160	209	20,565
Medium agent workflow	160	209	42,793
Large repeated audit	160	209	75,965
Sub-agent handoff example	160	209	54,476

Telemetry smoke

• Telemetry mode tested: summary
• Events captured: process_start, initialize, tools_list, tool_call, stdin_eof, process_exit
• Stores memory content: false

Interpretation

These numbers are not a universal guarantee. They show reproducible behavior on public synthetic fixtures. The expected value is highest when an agent would otherwise resend repeated project context across sessions.

This site is open source. Improve this page.