Skip to content

optimize-cli.md

Latest commit

 

History

History
272 lines (194 loc) · 7.53 KB

optimize-cli.md

File metadata and controls

272 lines (194 loc) · 7.53 KB
Raw

Optimize CLI Reference — v3.6

SuperLocalMemory V4 Documentation https://superlocalmemory.com | Part of Qualixar

Complete reference for all SLM v3.6 Optimize CLI commands.

Overview

SLM v3.6 adds 6 new top-level commands under slm:

slm optimize   ...    # Master Optimize module control
slm cache      ...    # Cache sub-control
slm compress   ...    # Compression sub-control
slm proxy      ...    # Start/proxy control
slm wrap       ...    # Agent activation
slm help-optimize     # Full developer reference

All commands accept --json for machine-readable output.

slm optimize — Master Module Control

slm optimize status

Show all Optimize settings — cache state, compression mode, proxy status, config version.

slm optimize status
slm optimize status --json

Example output:

Optimize: ON
  Cache:     enabled  (exact: 86400s TTL, semantic: OFF)
  Compress:  enabled  (Mode: safe, Prose (Layer 2): OFF)
  Proxy:     not running
  Config:    ~/.superlocalmemory/optimize.json  (version 61)

slm optimize on

Enable cache + compress (does NOT start the proxy).

slm optimize on

slm optimize off

Disable all Optimize features. Proxy (if running) passes through calls unchanged.

slm optimize off

slm optimize savings

Token/cost savings report from metrics counters.

slm optimize savings                    # Last 7 days
slm optimize savings --since 30         # Last 30 days
slm optimize savings --provider anthropic  # Filter by provider
slm optimize savings --json

Example output:

Savings (last 7 days):
  Exact cache hits:        43   (127,580 input tokens saved)
  Semantic cache hits:      0
  Tokens saved (total): 153,096
  Estimated savings:  ~$2.2964 (at $3.00/M tokens)
  Pricing date: 2026-06-07

slm cache — Cache Control

slm cache status

Entry count, DB size, TTLs, hit rate.

slm cache status                         # Default tenant
slm cache status --tenant myproject      # Specific tenant
slm cache status --json

Example output:

Cache status:
  Entries (exact):   127  (not expired)
  Semantic index:    OFF
  DB size:           2.4 MB  (~/.superlocalmemory/llmcache.db)
  TTL (exact):       86400s
  TTL (semantic):    3600s
  Hits:              43
  Misses:            89
  Hit rate:          32.6%

slm cache clear

Delete all entries for a tenant.

slm cache clear                          # Clear default tenant
slm cache clear --tenant myproject       # Clear specific tenant
slm cache clear --json

slm cache invalidate --tag <t>

Delete entries whose tag array contains the specified tag.

slm cache invalidate --tag "project:slm"
slm cache invalidate --tag "user:alice" --json

slm cache ttl

Set TTL values for cache entries.

slm cache ttl --set 3600           # 1 hour exact cache
slm cache ttl --semantic 7200      # 2 hour semantic cache
slm cache ttl --set 86400 --semantic 3600  # Both at once

Defaults: exact=86400s (24h), semantic=3600s (1h).

slm cache semantic on|off

Enable or disable semantic cache lookup (vCache).

slm cache semantic on
slm cache semantic off

Note: requires embedding model (~500MB). Run slm warmup if not already done.

slm compress — Compression Control

v3.6.10: the old slm compress code|ccr|align subcommands are removed — the homegrown extractive JSON/code compressor was lossy and has been deleted. Compression is now status | mode | prose. Running a removed subcommand prints a short notice explaining the change.

slm compress status

Show compression mode and prose state.

slm compress status
slm compress status --json

Example output:

Compression status:
  Enabled:  yes
  Mode:     safe        (lossless normalization only)
  Prose:    OFF         (LLMLingua-2, aggressive mode only)

slm compress mode safe|aggressive

Set compression aggressiveness.

slm compress mode safe          # Default — LOSSLESS (whitespace/JSON normalization only; code untouched)
slm compress mode aggressive    # ⚠ Adds LLMLingua-2 prose compression (prose only)

Safe mode (default): lossless — whitespace collapse + compact-JSON re-serialization only. No content is removed; code, numbers, structured data, and the current turn are never touched. Production-safe for all use cases, and keeps the provider prefix-cache stable.

Aggressive mode: additionally enables LLMLingua-2 prose compression on prose blocks only — may omit nuance, hedges, or low-salience wording. Shows a warning before activating. Never compresses code, math, structured data, instructions, or the latest user turn.

slm compress prose on|off

Enable/disable LLMLingua-2 prose compression (takes effect in aggressive mode).

slm compress prose on
slm compress prose off

slm proxy — Interception Proxy

Start or report the optimization proxy. Proxy intercepts LLM API calls, applies cache lookup and compression before forwarding.

slm proxy                              # Default port 8765
slm proxy --port 8080                  # Custom port
slm proxy --provider anthropic         # Provider-specific surface
slm proxy --no-compress                # Cache only, no compression
slm proxy --semantic                   # Enable semantic cache on proxy
slm proxy --json

Endpoints served on proxy:

  • POST /v1/messages — Anthropic Messages API
  • POST /v1/chat/completions — OpenAI Chat API
  • POST /v1beta/models/* — Gemini native API
  • Compatible with all OpenAI-compatible clients

slm wrap — Agent Activation

Activate an agent through the SLM proxy. One command: starts proxy + sets environment + launches agent.

slm wrap claude                       # Claude Code (recommended)
slm wrap claude -- my prompt          # Pass args to agent
slm wrap cursor                       # Cursor
slm wrap aider -- --model gpt-4       # Aider with model
slm wrap antigravity                  # Antigravity
slm wrap --list                       # List registered agents
slm wrap --persistent                 # Write to agent's config file
slm wrap --dry-run                    # Preview without executing

Supported agents: claude, claude-settings, antigravity, cline, opencode, cursor, aider, codex, copilot, generic

SDK adapters (Python) — manual wrap:

from superlocalmemory.optimize.adapters.openai_adapter import withSLM
from openai import OpenAI
client = withSLM(OpenAI())        # Same interface — zero API change

from superlocalmemory.optimize.adapters.anthropic_adapter import withSLM
from anthropic import Anthropic
client = withSLM(Anthropic())

slm help-optimize — Developer Reference

Full reference with per-agent setup recipes and safety notes.

slm help-optimize                     # Full reference
slm help-optimize cache               # Cache subcommand reference
slm help-optimize compress            # Compress + safety warning
slm help-optimize agents              # Per-agent setup recipes
slm help-optimize safety              # Compression safety warning only

Related Pages

Part of Qualixar | Created by Varun Pratap Bhardwaj | AI Reliability Engineering